International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 07 | Jul 2026
p-ISSN: 2395-0072
www.irjet.net
A Hybrid Task-Oriented Artificial Intelligence Framework for Video Compression with Embedded Object-Recognition Parameters: An Approach Beyond H.265+ Semantic-Aware Video Compression with Object-Parameter Sidestream (SAVC-OPS) Samir Khanji Informatics and Communications Department, Faculty of Engineering, Ebla Private University, Idlib, Syria ---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - The exponential growth of surveillance and
Key Words: artificial intelligence video compression; H.265+; HEVC; Video Coding for Machines; object recognition; feature coding; semantic storage; analytical simulation; surveillance systems.
machine-vision workloads has made bitrate and storage efficiency a first-order systems constraint. Vendor-optimized codecs such as H.265+ improve upon High Efficiency Video Coding (HEVC) for static-heavy scenes through background modeling, noise suppression, and long-term average bitrate control, while newer standards such as VVC further raise compression efficiency for human viewing [2]. Nevertheless, these codecs remain primarily oriented toward human perception and do not, by themselves, define a standardized persistent object-recognition parameter track inside the compressed representation (some NVR products may store analytics metadata outside the codec). Consequently, archives must often be re-analyzed with costly neural inference whenever semantic queries are required. Related research on Video Coding for Machines (VCM) and Feature Coding for Machines (FCM) reframes compression around task utility rather than pixel fidelity alone [1], [8], [16]. This paper proposes SAVC-OPS (Semantic-Aware Video Compression with Object-Parameter Side-stream), a hybrid framework that combines: (i) task-aware semantic bit allocation based on detected regions of interest [1], [9]; (ii) a temporally compressed object-parameter side-stream that stores bounding boxes, class labels, confidences, track identifiers, and optional embeddings; and (iii) an optional neural featurecoding path for machine-primary consumption [8], [15]. Using a calibrated illustrative analytical simulation (not measured bitstreams), mean bitrate decreases from 3.17 Mbps (H.265) and 1.33 Mbps (H.265+) to 1.17 Mbps (SAVC-A) and 0.74 Mbps (SAVC-B). Thirty-day storage per camera falls from about 421 GB under H.265+ to 370 GB (SAVC-A) and 234 GB (SAVC-B). Under a comparable storage budget, illustrative mAP@0.5 improves from 0.77 to 0.85–0.88, while semantic query latency can fall by roughly 400× through OPS indexing (hardwaredependent assumption). Five incremental development proposals are quantified with respect to storage, task accuracy, retrieval efficiency, edge cost, and privacy. The results indicate that moving from scene-aware compression to task-aware, semantically annotated compression is a practical path beyond H.265+ for analytics-centric video systems.
© 2026, IRJET
|
Impact Factor value: 8.226
1. INTRODUCTION 1.1 Motivation and Context Video is the dominant contributor to digital storage and network traffic in contemporary infrastructure, particularly in video surveillance, smart-city sensing, autonomous platforms, and remote healthcare. The transition from 1080p to 4K/8K capture, together with denser camera deployments, multiplies both bandwidth demand and archival cost. Over the last decade, standardization and industrial engineering have therefore focused on improving compression efficiency while preserving acceptable visual quality for human operators. Modern standardized codecs continue to push rate– distortion efficiency for human viewing: Versatile Video Coding (VVC / H.266) approximately halves bitrate relative to HEVC for comparable quality in many conditions [2]. In parallel, surveillance vendors introduced proprietary HEVCoriented extensions commonly marketed as H.265+ (and closely related “smart” codecs). H.265+ is not an ITU-T/ISO international standard; it denotes vendor optimizations layered on HEVC. These extensions exploit properties typical of fixed-camera deployments: long quasi-static backgrounds, intermittent motion, and sensor noise that would otherwise consume bits without operational value. Reported industrial examples indicate that average bitrate under H.265+ may fall to roughly one-third of H.265 in low-activity office scenes, although the gain is strongly content dependent. Despite these advances, the consumer of video is changing. In an increasing fraction of deployments, the primary consumer is not a human viewer but a machinevision pipeline that performs detection, tracking, reidentification, counting, or anomaly analysis. Under this regime, peak signal-to-noise ratio (PSNR) and related
|
ISO 9001:2008 Certified Journal
|
Page 342
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 07 | Jul 2026
p-ISSN: 2395-0072
www.irjet.net
perceptual metrics are incomplete success criteria. What matters is the preservation of task-relevant information at the lowest feasible rate, together with the ability to retrieve semantic events from archives without repeatedly decoding and re-inferring every frame.
• Formalization. We define the side-stream syntax, temporal differential coding principles, and a multi-objective rate–task–perception loss suitable for mode selection (human, hybrid, machine). • Comparative evaluation protocol. We specify fair comparison axes against H.265 and H.265+: bitrate by scene class, thirty-day storage, rate–task curves, perceptual metrics, detection/tracking metrics, and semantic query latency.
1.2 Limitations of H.265+-Centric Pipelines H.265+ improves the human-oriented rate–distortion tradeoff for surveillance content, yet three structural limitations remain.
• Analytical simulation study. We report quantified comparisons and seven publication-ready figures based on a calibrated model, clarifying where gains over H.265+ are expected to be large or modest.
First, bit allocation is driven by background/motion/noise cues rather than by semantic task relevance. Regions that are visually static may still contain small or slowly moving objects that are critical for detection. Conversely, textured background motion may attract bits that contribute little to downstream accuracy.
• Development roadmap. We decompose the framework into five proposals (P1–P5), each with advantages, challenges, and simulated outcomes, and provide a staged adoption plan.
Second, object-recognition parameters if computed at all on the edge are typically treated as ephemeral analytics outputs. They are not stored as a compact, synchronized component of the compressed archive. Later investigations therefore re-run detectors and trackers on reconstructed pixels, multiplying compute cost and delaying response.
1.5 Paper Organization Section 2 reviews related work. Section 3 states the problem, objectives, and hypotheses. Section 4 presents the SAVC-OPS methodology. Section 5 details the simulation setup. Section 6 reports statistical comparisons and figures. Section 7 develops proposals with advantages and results. Section 8 discusses implications and limitations. Section 9 concludes the paper.
Third, emerging machine-to-machine scenarios may not require a high-fidelity pixel archive at all. Feature Coding for Machines (FCM) and related approaches suggest that intermediate neural representations can be compressed more efficiently for remote inference [1], [8], [16]. H.265+ does not natively expose such a path.
2. RELATED WORKS
1.3 Research Gap
2.1 From H.264 to HEVC and Surveillance-Oriented H.265+
The literature now contains strong strands on learned codecs [3], [4], [5], [17], [18], Video Coding for Machines (VCM) [1], [6], [7], ROI retargeting, reinforcement-learning bit control [9], and feature coding [8], [15], [16]. However, practical surveillance stacks still largely separate (a) pixel compression optimized for human viewing from (b) analytics executed as a post-process. A unified representation that jointly optimizes storage, task accuracy, and semantic retrievability while remaining deployable on HEVC-compatible infrastructure remains insufficiently articulated for systems that currently rely on H.265+.
HEVC and its successor VVC introduced successively stronger coding tools (flexible partitions, refined prediction, and improved entropy coding), with VVC targeting roughly a further ~50% bitrate reduction over HEVC under matched quality [2]. Industrial H.265+ extensions typically emphasize three tools tailored to surveillance: (i) background-modelbased prediction, (ii) background noise suppression to avoid spending bits on non-informative sensor fluctuations, and (iii) long-term average bitrate control that reallocates bits across busy and idle periods of a day. These tools can dramatically reduce storage for low-motion cameras. Their optimization target, however, remains a human-viewable reconstruction rather than a machine task metric such as mean average precision (mAP) or multi-object tracking accuracy (MOTA) [1], [13].
1.4 Contributions This paper makes the following contributions: • Framework design. We introduce SAVC-OPS, a hybrid architecture that couples semantic bit allocation with a persistent object-parameter side-stream and an optional feature-coding track.
© 2026, IRJET
|
Impact Factor value: 8.226
2.2 Learning-Based Image and Video Compression Deep learning has produced end-to-end nonlinear transform codecs with learned entropy models that are competitive with classical still-image standards at selected
|
ISO 9001:2008 Certified Journal
|
Page 343
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 07 | Jul 2026
p-ISSN: 2395-0072
www.irjet.net
operating points [5]. Extending these ideas to video has progressed through predictive/conditional frameworks such as DCVC [3], transformer-based temporal entropy models such as VCT [4], and content-adaptive neural codecs [17], alongside broader surveys of intelligent video coding [18]. Real-time complexity and edge constraints nonetheless remain challenging. Hybrid designs are therefore attractive: a conventional HEVC/VVC core is retained for hardware decode compatibility, while neural modules guide preprocessing, mode decisions, or post-processing. SAVCOPS follows this hybrid philosophy for near-term deployability.
statistics drift, but they also introduce training cost and domain-transfer challenges. We therefore treat RL-based allocation as an incremental proposal (P4) rather than a mandatory component of the baseline stack.
2.3 Video Coding for Machines (VCM)
Approach
VCM reframes compression as a rate–task problem [1]. Instead of minimizing pixel distortion alone, encoders allocate quality to regions and features that preserve automated analysis. Representative techniques include ROIaware quantization, saliency- or detector-guided CTU importance, and multi-scale feature pipelines evaluated under MPEG VCM common test conditions [6], [7]. Recent peer-reviewed studies report large BD-rate gains for machine tasks under feature/video VCM configurations [6], and standardization surveys summarize use cases, metrics, and proposal trends [7], [8]. These results motivate semantic bit allocation in SAVC-OPS, but most VCM pipelines still treat ROI/feature metadata as encoder side information rather than as a durable archival track for semantic search.
H.265/HEVC Human R-D (pixels) No No / VVC [2] Human Background/ H.265+ No No (surveillance) motion/noise Typical VCM Machine (+ Transient / ROI tools [1], Semantic ROI Partial human) limited [6], [7] FCM [8], Feature rate– Features, Machine Yes [15], [16] accuracy not boxes Human and/ Semantic map Yes (OPS SAVC-OPS Optional or machine + task loss track)
2.6 Positioning of This Work Table -1 summarizes the conceptual positioning relative to VCM/FCM surveys and codecs [1], [2], [7], [8]: Table -1: Simulation Parameters.
3.1 Problem Statement
FCM compresses intermediate neural activations rather than (or in addition to) pixels [8], [15], [16]. This reduces bandwidth for remote inference, supports compute offloading, and can improve privacy by avoiding transmission of the full scene. Analyses of machine-tomachine coding indicate that conventional inner codecs such as HEVC/VVC remain highly effective for feature tensors with only modest task BD-rate differences between HEVC and VVC in reported FCM experiments which is encouraging for hardware reuse [8]. Learnable multi-scale feature compression further improves the rate–task curve for VCMstyle pipelines [6], [15]. SAVC-OPS incorporates an optional feature track for machine-primary modes while retaining a pixel track for auditability when required.
Let denote a video sequence. A conventional surveillance encoder produces a bitstream minimizing a perceptual rate–distortion objective. An analytics module may later compute object sets from the reconstructed frames . The operational cost of an archive is then not only but also the repeated inference cost whenever semantic queries are issued.
that jointly reduces storage relative to H.265+, preserves or improves task metrics, and makes (object parameters) and optionally (features) first-class, queryable components of the archive.
2.5 Task-Aware Control and Reinforcement Learning
3.2 Research Question
Recent work explores reinforcement learning agents that select block-level quantization actions to optimize downstream task performance under rate constraints. In particular, RL-RC-DoT demonstrates task-aware macroblock QP control on top of a standard encoder without requiring the task network at inference time [9]. Such controllers can outperform static ROI heuristics when scene
|
Impact Factor value: 8.226
Object MachineBit decision parameters primary unit persisted? path
3. PROBLEM STATEMENT, OBJECTIVES, AND HYPOTHESES
2.4 Feature Coding for Machines (FCM)
© 2026, IRJET
Primary consumer
How can an AI-assisted compression framework surpass H.265+ in storage efficiency and task utility for analyticscentric video, while embedding object-recognition parameters in the compressed representation?
|
ISO 9001:2008 Certified Journal
|
Page 344
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 07 | Jul 2026
p-ISSN: 2395-0072
www.irjet.net
3.3 Objectives
4.2 Perception Unit and Object Parameterization
• Design a hybrid architecture compatible with HEVCclass cores. • Specify a compact, temporally coded object-parameter side-stream. • Define a multi-axis evaluation protocol (rate, perception, task, retrieval). • Quantify expected gains via analytical simulation and structured proposals.
For each frame
, the perception unit outputs
where is a normalized bounding box, a class index, a confidence score, a temporal track identifier, and an optional embedding (e.g., for re-identification). Lightweight real-time detectors such as YOLOv8-class models [11] and association trackers such as ByteTrack [10] are appropriate on edge hardware.
3.4 Hypotheses • H1: Under fixed target detection accuracy, semantic bit allocation reduces total size relative to H.265+ for medium and crowded scenes. • H2: The storage overhead of OPS is small, yet semantic retrieval latency decreases by orders of magnitude. • H3: For machine-primary consumption, a featureaugmented path outperforms a pixel-only path on the rate–task curve.
4.3 Semantic Importance Map A dense map is constructed by rasterizing boxes/masks, dilating them by a motion-predicted margin, and assigning a low residual weight to long-term background. Formally,
4. PROPOSED METHODOLOGY: THE SAVC-OPS FRAMEWORK
where denotes predicted motion and is a deterministic fusion operator. Unlike H.265+ background models, is class- and task-aware: a small pedestrian in a textured plaza can receive higher priority than large foliage motion.
4.1 Architectural Overview Fig -1 presents the architecture. Incoming frames are processed by a perception unit (detector and tracker). The resulting object set induces a semantic importance map that drives bit allocation in a video core codec. In parallel, object parameters are differentially coded into an OPS track. Optionally, intermediate features are coded into a feature track. All tracks share a synchronization timeline and may be accompanied by a lightweight semantic index.
4.4 Semantic Bit Allocation Given , the allocator produces codec control variables. Two practical realizations are supported: • ROI-QP control. Coding tree units overlapping highimportance regions receive ; background regions receive . • ROI retargeting. The frame is geometrically warped so that ROI regions occupy more pixels before encoding; inverse geometry is signaled as metadata for reconstruction. A hybrid policy may apply mild retargeting followed by residual QP shaping. The video core may be x265/HEVC hardware, VVC [2], or a learned codec [3], [4], provided decode compatibility constraints of the deployment are respected.
4.5 Object-Parameter Side-stream (OPS)
Fig -1: . SAVC-OPS architecture: perception, semantic mapping, bit allocation, video core, OPS encoder, optional feature encoder, and multiplexed container.
© 2026, IRJET
|
Impact Factor value: 8.226
OPS persists recognition outputs as a compressed archival track rather than discarding them after ROI decisions. Temporal coding proceeds as follows:
|
ISO 9001:2008 Certified Journal
|
Page 345
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 07 | Jul 2026
p-ISSN: 2395-0072
www.irjet.net
4.8 Multi-Objective Training / Control Loss
• Associate objects across time via . • Predict box trajectories (constant velocity or Kalmanstyle predictors).
Allocation parameters are selected to minimize
•Entropy-code residuals , class persistence flags, and confidence deltas. • Emit birth/death events instead of repeating full state every frame. • Optionally insert semantic keyframes every for random access.
where are track rates, penalizes task degradation (or uses a differentiable proxy), and is an optional perceptual term for human-facing modes. Weight schedules differ by mode: human archives raise ; machine archives raise and .
frames
In the simulation, OPS contributes approximately 2.1% of the video-track size under SAVC-A small relative to its retrieval benefit. A logical container is:
4.9 Encoding Algorithm
SAVC-OPS Container
video_track
# HEVC/VVC/AI bitstream
ops_track
# compressed object parameters
Algorithm 1: SAVC-OPS Encode Input: video V, mode ∈ {human, hybrid, machine} Output: container C for each frame x_t in V do O_t ← DetectAndTrack(x_t) S_t ← BuildSemanticMap(O_t, history) if mode ≠ machine_only then x_t ← PreprocessROI(x_t, S_t) B_v ← VideoEncode(x_t, S_t) else B_v ← PreviewEncode(x_t) or ∅ B_o ← OPSEncode(O_t, O_{t-1}) if mode ∈ {hybrid, machine} then f_t ← ExtractFeatures(x_t) B_f ← FeatureEncode(f_t, f_{t-1}) else B_f ← ∅ Mux(C, B_v, B_o, B_f, t) return C
feature_track? # optional neural features sync_timeline # PTS/frame indices index_sidecar # optional (class/time/camera)
inverted
index
4.6 Optional Feature-Coding Path When the consumer is primarily a machine, intermediate backbone features are quantized and temporally coded:
Downstream heads (detection/tracking) operate on reconstructed features, enabling aggressive reduction of the pixel track to a low-rate preview or even an empty track in pure machine mode. This path aligns with FCM concepts [8], [15], [16] while remaining optional in hybrid deployments that require human audit frames.
4.10 Why SAVC-OPS Can Surpass H.265+ Dimension Decision unit
4.7 Operating Modes Mode
Active tracks
Intended use
SAVC-A
Video + OPS
Hybrid archive: human review + semantic search
SAVC-B
Low-rate video/ Machine-heavy analytics preview + OPS + with limited auditability features
Humanfallback
Video (+ transient Compatibility mode ROI) close to classical codecs
© 2026, IRJET
|
Impact Factor value: 8.226
H.265+ SAVC-OPS Background / Task semantics + motion motion / noise
Visual quality per Quality + task + retrieval bit cost Object parameters Not persisted Compressed OPS track Machine-only path Weak Optional FCM-like track Success metric
Re-analysis
Full decode inference
+
Often OPS query only
Superiority is not claimed for every scene. Gains concentrate where semantics are extractable and archives are repeatedly queried.
|
ISO 9001:2008 Certified Journal
|
Page 346
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 07 | Jul 2026
p-ISSN: 2395-0072
www.irjet.net
5. ANALYTICAL SIMULATION SETUP
OPS coding Temporal residuals + entropy coding Features (SAVC-B) Intermediate backbone maps, 8-bit quantization
5.1 Reference Scenario
5.5 Scope of Validity
Parameter Resolution / frame rate Archive horizon Scene classes Baselines Proposed H.265+ calibration
Value 4 MP @ 25 fps, continuous 30 days per camera Low motion; medium; crowded H.265; H.265+ SAVC-A; SAVC-B Vendor white-paper trends (larger savings in low motion; smaller in crowded scenes) Semantic gain calibration VCM/FCM rate–task trends [1], [6], [7], [8] Metrics Mbps; GB/30d; PSNR/SSIM/VMAF; structure– texture IQA [12]; mAP; MOTA/IDF1 [13], [14]; query latency
The model captures first-order trade-offs suitable for design and roadmap decisions. It does not replace codec conformance testing, vendor bitstream measurement, or dataset-specific training of detectors. Section 8 discusses threats to validity.
6. RESULTS: STATISTICS, COMPARISONS, AND FIGURES 6.1 Bitrate by Scene Type Table- 2. Average bitrate (Mbps) by scene class.
5.2 Storage Conversion Thirty-day storage in gigabytes is approximated by
Scene
H.265
Low motion Medium Crowded Overall mean
2.40 3.00 4.10
SAVC-A H.265+ SAVC-A SAVC-B saving vs H.265+ 0.80 0.85 0.48 −6.3% 1.15 1.05 0.68 8.7% 2.05 1.60 1.05 22.0%
3.17
1.33
1.17
0.74
12.0%
Using this factor, 1.33 Mbps corresponds to approximately 421 GB per camera per 30 days. All storage tables in Section 6 are computed from the same factor to keep bitrate and storage mutually consistent. 5.3 Fairness Axes Comparisons are reported under multiple matched conditions:
• Bitrate by scene class (content-aware). • Monthly storage composition (video / OPS / features). • Rate–task curves (mAP@0.5 versus Mbps). • Perceptual metrics at a fixed 1.5 Mbps point. • Detection/tracking metrics at a matched ~1.7 Mbps budget. • Semantic query latency on a one-hour archive.
Fig- 2. Bitrate comparison across scene classes. SAVC gains are modest in very static scenes where H.265+ is already strong and become pronounced in medium and crowded scenes.
5.4 Implementation Choices Assumed by the Model Component Detector Tracker Video core ROI QP deltas
© 2026, IRJET
H.265+ captures much of the available redundancy when backgrounds dominate. Semantic allocation adds the most value when many task-critical objects compete with complex backgrounds.
Assumed practical choice YOLO-class nano/small [11] ByteTrack-style association [10] HEVC (x265-class / hardware); VVC optional [2] (ROI), (background)
|
Impact Factor value: 8.226
|
ISO 9001:2008 Certified Journal
|
Page 347
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 07 | Jul 2026
p-ISSN: 2395-0072
www.irjet.net
6.2 Thirty-Day Storage Composition Table- 3. Estimated storage (GB / 30 days / camera). Method Video OPS Features Total Saving vs H.265+ H.265
1003.0 0
0
1003.0 −138.4% (larger)
H.265+ 420.8 0
0
420.8 reference
SAVC-A 362.4 7.8 0
370.2 12.0%
SAVC-B 140.0 7.8 86.3
234.1 44.4% Fig- 4. Rate–task curves. At matched bitrates, SAVC modes preserve higher detection accuracy because bits are concentrated on semantically critical regions (and on features in SAVC-B). An approximate task-equivalent comparison in the model: reaching mAP@0.5 ≈ 0.83 requires about 2.5 Mbps for H.265+ but about 1.5 Mbps for SAVC-A, i.e., roughly 40% illustrative rate reduction at equal task accuracy. These curves are design aids, not codec conformance results.
6.4 Perceptual Quality at Fixed 1.5 Mbps Table 5. Human perceptual metrics and task accuracy at 1.5 Mbps.
Fig- 3. Stacked storage composition. OPS adds a small absolute overhead (~7.8 GB/month) while enabling semantic indexing. In SAVC-B, feature storage partially replaces pixel storage.
6.3 Rate–Task Curves Table -4. mAP@0.5 as a function of bitrate (Mbps). Mbps
H.265
H.265+
SAVC-A
Method PSNR (dB) ↑
SSIM ↑
VMAF ↑
mAP@0.5 ↑
H.265
36.8
0.941
78.2
0.76
H.265+
37.0
0.943
78.8
0.75
SAVC-A
36.1
0.936
76.8
0.83
SAVC-B
31.5
0.888
59.0
0.86
SAVC-B
0.5
0.52
0.51
0.61
0.68
1.0
0.68
0.67
0.76
0.80
1.5
0.76
0.75
0.83
0.86
2.0
0.81
0.80
0.86
0.885
2.5
0.84
0.83
0.88
0.90
3.0
0.86
0.85
0.89
0.905
H.265+ can remain slightly superior on global perceptual scores because background cleaning improves average fidelity. SAVC-A/B intentionally reallocate quality toward objects, which may lower global PSNR/VMAF while raising mAP. This dissociation is central: analytics systems should not optimize PSNR alone, consistent with modern structure/texture-aware IQA arguments and with VCM’s rate–task objective [1], [12].
6.5 Detection and Tracking under a Matched Storage Budget Tracking-oriented evaluation follows the spirit of crowded-scene MOT benchmarks [13], [14], while detection uses mAP@0.5 as in common detector reporting [11].
© 2026, IRJET
|
Impact Factor value: 8.226
|
ISO 9001:2008 Certified Journal
|
Page 348
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 07 | Jul 2026
p-ISSN: 2395-0072
www.irjet.net
Table- 6. Task metrics at an approximately matched ~1.7 Mbps budget.
6.7 Aggregate Assessment
Method mAP@0.5 ↑ mAP@0.5:0.95 ↑ MOTA ↑ IDF1 ↑
Table - 9. Headline comparison: H.265+ versus best SAVC mode.
H.265
Summary
0.78
0.51
0.62
0.66
H.265+ 0.77
0.50
0.61
0.65
Indicator
H.265+
SAVC-A 0.85
0.57
0.70
0.73
Mean bitrate
1.33 Mbps
SAVC-B 0.88
0.60
0.73
0.76
30-day storage 421 GB mAP@0.5 @ matched 0.77 budget Class-query latency 186 s
Table- 7. Absolute gains versus H.265+ (percentage points). Metric
Δ SAVC-A
Δ SAVC-B
mAP@0.5
+8.0
+11.0
mAP@0.5:0.95
+7.0
+10.0
MOTA
+9.0
+12.0
IDF1
+8.0
+11.0
OPS overhead
0%
and
Hypothesis
Best SAVC Outcome 0.74 Mbps −44% (B) 234 GB (B) −44% 0.88 (B) 0.42 s ≈2.1% video
+11 pp ≈443× faster of operationally acceptable
Retrieval latencies depend on hardware-dependent latency assumptions (decode + detector throughput versus indexed side-stream lookup). Within the simulation: H1 is supported for medium/crowded scenes; H2 is strongly supported; H3 is supported by SAVC-B dominance on the rate–task and machine-storage axes.
6.6 Semantic Retrieval Latency Table- 8. Query latency on a one-hour archive (seconds).
7. DEVELOPMENT PROPOSALS: ADVANTAGES AND RESULTS
H.265+ (re- SAVC-OPS Speedup inference) (OPS query) 186 0.42 ≈443× 198 0.38 ≈521×
The framework is decomposed into five adoptable proposals. Fig- 6 and Fig- 7 summarize comparative gains and multi-criteria scores.
Query type Class filter Track-ID trajectory Object-density window
210
0.55
≈382×
Figure 6. Relative gains of proposals P1–P5 versus H.265+ in bitrate saving, mAP gain at 1.5 Mbps, and retrieval speedup (normalized display).
Fig- 5. Retrieval latency comparison. Persistent OPS transforms repeated GPU inference into indexed sidestream queries.
© 2026, IRJET
|
Impact Factor value: 8.226
|
ISO 9001:2008 Certified Journal
|
Page 349
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 07 | Jul 2026
p-ISSN: 2395-0072
www.irjet.net
• Avoids repeated full-archive inference. • Supports multi-camera event correlation through side indices. Challenges. • Creates a sensitive personal-data artifact requiring encryption and retention policy. • Requires robust clock/frame synchronization. Simulated results. • Storage overhead ≈ 2.1% of video size (~7.8 GB/30 days in the reference scenario). • Retrieval speedup ≈ 380–520× (illustrative). • Minor task-side benefit (+0.5 pp) via more reliable offline association workflows.
Figure 7. Multi-criteria scores (0–100) for storage efficiency, task accuracy, retrieval, edge cost, and privacy.
7.1 Proposal P1 - Adaptive Semantic ROI-QP
7.3 Proposal P3 - Optional Feature-Coding Path (FCM-like)
Concept. Replace purely background/motion QP logic with detectordriven importance maps and CTU-level .
This proposal follows the FCM/split-inference direction of MPEG-AI and recent machine-to-machine coding analyses [8], [15], [16].
Advantages.
Concept.
• High compatibility with existing HEVC/VVC hardware encoders [2], [9]. • Limited incremental complexity if analytics already run on the edge. • Immediate improvement of the rate–task operating curve.
For machine-primary modes, compress intermediate features and reduce reliance on high-rate pixels. Advantages. • Largest analytical storage reduction. • Improved privacy relative to full-fidelity video export. • Natural fit for split inference between edge and cloud.
Challenges. • Sensitivity to detector errors (false positives inflate ROI area). • Possible reduction of global PSNR even when task metrics improve.
Challenges. • Weak human preview quality unless a parallel low-rate video track is kept. • Feature-format coupling to model families complicates long-term archive readability.
Simulated results. • Bitrate saving ≈ 12% versus H.265+ at approximately matched task accuracy. • mAP@0.5 gain ≈ +6.0 pp at 1.5 Mbps. • Retrieval speedup ≈ 1× (no persistent index by itself).
Simulated results. • Bitrate saving up to ≈ 40% versus H.265+ in machine mode. • mAP@0.5 gain ≈ +8.0 pp at 1.5 Mbps. • Retrieval/analysis speedup ≈ 12× versus full pixel reinference when features and OPS are indexed.
7.2 Proposal P2 - Persistent OPS Side-stream Concept. Store temporally coded object parameters as a durable track synchronized with video.
7.4 Proposal P4 - Reinforcement-Learning BitAllocation Controller
Advantages.
P4 is inspired by block-level RL controllers for task-aware QP assignment on standard encoders [9].
• Enables interactive semantic search and rapid audit.
© 2026, IRJET
|
Impact Factor value: 8.226
|
ISO 9001:2008 Certified Journal
|
Page 350
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 07 | Jul 2026
p-ISSN: 2395-0072
www.irjet.net
7.6 Consolidated Proposal Table
Concept. Train an RL policy to select block/ROI encoding actions that maximize expected future task reward under a rate constraint.
Table- 10. Development proposals, advantages, and simulated outcomes.
Advantages.
Principal ID Core idea advantage
• Adapts to nonstationary scenes better than fixed ROI heuristics. • Can transfer across related tasks when rewards are carefully designed.
Bitrate saving Δ mAP @ Retrieval vs 1.5 Mbps speedup H.265+
Semantic Easy hardware 12% ROI-QP fit Persistent Instant P2 −2% OPS semantic query Feature Max analytical P3 40% coding compression RL Scene-adaptive P4 18% allocation control P5 Full stack Best balanced 28% P1
Challenges. • Training is expensive (encode + task evaluation in the loop). • Edge deployment may require distillation to a compact policy network.
+6.0 pp
~1×
+0.5 pp
~450×
+8.0 pp
~12×
+7.0 pp
~1.2×
+9.0 pp
~480×
P2 may slightly increase bitrate (OPS overhead) while sharply reducing retrieval/analysis cost; its value is operational rather than pure compression.
Simulated results. • Bitrate saving ≈ 18%. • mAP@0.5 gain ≈ +7.0 pp at 1.5 Mbps. • Retrieval impact modest (~1.2×) via more stable tracks.
7.7 Staged Adoption Roadmap Table - 11. Recommended deployment stages.
7.5 Proposal P5 - Full SAVC-OPS Stack
Stage
Concept. 1-Fast path
Integrate P1–P4 according to operating mode (A/B), with shared synchronization and security controls.
2-Adaptive control
Advantages. • Best overall trade-off among storage, accuracy, and retrieval. • Provides a coherent product architecture rather than isolated tools.
3-Machine mode
Horizon
Proposals Expected outcome Modest rate benefit in active scenes + 1–2 months P1 + P2 semantic queryability 3–5 months + P4
Additional rate– task improvement under scene drift
6–9 months + P3
Larger analytical storage reduction with preserved task accuracy
Challenges.
8. DISCUSSION
• Highest integration complexity and MLOps burden. • Requires governance for semantic data lifecycle.
8.1 When Does SAVC-OPS Beat H.265+?
Simulated results. In low-motion scenes, H.265+ already removes much temporal redundancy; SAVC-A bitrate gains are small, and the decisive benefit is retrieval via OPS. In medium and crowded scenes, semantic allocation yields clearer rate and task gains because background-centric heuristics become less aligned with object density. In archives that are frequently re-analyzed, OPS can dominate the business case even when compression gains are moderate.
• Mean bitrate saving ≈ 28% versus H.265+. • mAP@0.5 gain ≈ +9.0 pp at 1.5 Mbps. • Retrieval speedup ≈ 480× (illustrative). • Thirty-day storage ≈ 234 GB versus 421 GB for H.265+.
© 2026, IRJET
|
Impact Factor value: 8.226
|
ISO 9001:2008 Certified Journal
|
Page 351
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 07 | Jul 2026
p-ISSN: 2395-0072
www.irjet.net
8.2 Perception–Task Trade-off
9. CONCLUSION AND FUTURE WORK
Table-4 shows that improving mAP can coincide with lower global PSNR/VMAF. System designers must explicitly choose the objective. For evidence rooms and human investigation, SAVC-A with constrained background QP is preferable. For large-scale automated mining, SAVC-B is preferable.
This paper presented SAVC-OPS, a hybrid AI-assisted video compression framework designed for analytics-centric archives that currently rely on H.265/H.265+. By combining semantic bit allocation, a persistent object-parameter sidestream, and an optional feature-coding path, the framework targets not only bitrate reduction but also task preservation and semantic retrievability.
8.3 Edge Cost Accounting
Within a calibrated analytical simulation for 4 MP / 25 fps continuous capture:
If cameras or recorders already run detection for alerts, SAVC-OPS largely reuses those outputs. The incremental cost is ROI control, OPS entropy coding, and optional feature coding. This amortization argument is essential: comparing SAVC against a codec that performs no analytics understates real deployments where analytics exist but are discarded after alerting.
• Illustrative mean bitrate falls from 1.33 Mbps (H.265+) to 1.17 Mbps (SAVC-A) and 0.74 Mbps (SAVCB). • Thirty-day storage decreases by approximately 12% (SAVC-A) to 44% (SAVC-B) in the model. • At a matched storage budget, illustrative mAP@0.5 improves by about +8 to +11 percentage points. • Semantic queries can accelerate by roughly 400–500× through OPS under stated latency assumptions. • Five development proposals provide a practical adoption ladder, with the full stack (P5) offering the best balanced modeled gain.
8.4 Privacy, Security, and Governance OPS and embeddings can encode movement patterns and identity cues. Recommended controls include: separate encryption keys for OPS/feature tracks; role-based access stricter than for video; retention schedules with semantic deletion; and minimization (e.g., omit face embeddings unless legally justified). Paradoxically, FCM-like modes may reduce exposure of raw pixels while still requiring careful feature governance.
Practical recommendation. Deploy P1 + P2 first for rapid value (semantic QP control and queryable object parameters), then add P4 and P3 as adaptive control and machine-primary modes mature.
8.5 Standardization Outlook
Future work includes bitstream-level experiments on public and proprietary surveillance sets, closed-loop training of allocation policies with task rewards [9], compact temporal coding of re-identification embeddings, and formal mapping onto emerging MPEG VCM/FCM syntaxes [7], [8], [16]. Field studies across seasonal lighting and multi-vendor H.265+ devices are required to convert the present design evidence into fully measured deployment claims.
Alignment with MPEG VCM/FCM increases long-term interoperability [7], [8], [16]. Near-term systems can ship private tracks inside MP4/ISOBMFF or parallel sidecar files with explicit synchronization, then migrate to standardized semantics as they mature. Keeping the video track HEVC/VVC-compatible preserves decoder ecosystems during transition [2], [8].
REFERENCES
8.6 Threats to Validity
1) Duan, L.-Y., Liu, J., Yang, W., Huang, T., & Gao, W. (2020). Video coding for machines: A paradigm of collaborative compression and intelligent analytics. IEEE Transactions on Image Processing, 29, 8680–8695. https://doi.org/10.1109/TIP.2020.3016485
• Calibration uncertainty. Industrial H.265+ claims vary by vendor and scene; local measurement is mandatory before procurement decisions. • Detector dependence. ROI and OPS quality inherit detector/tracker errors; confirmation mechanisms (periodic keyframe re-detection, human-in-the-loop correction) are needed for forensic use. • Model mismatch for features. Feature archives may become obsolete if downstream models change; versioned feature schemas and periodic re-extraction policies mitigate this risk. • Simulation abstraction. Interaction effects among QP, retargeting, and B-frames are simplified; bitstream-level experiments are the next validation step.
© 2026, IRJET
|
Impact Factor value: 8.226
2) Bross, B., Wang, Y.-K., Ye, Y., Liu, S., Chen, J., Sullivan, G. J., & Ohm, J.-R. (2021). Overview of the Versatile Video Coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10), 3736–3764. https://doi.org/10.1109/TCSVT.2021.3101953
|
ISO 9001:2008 Certified Journal
|
Page 352
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 07 | Jul 2026
p-ISSN: 2395-0072
www.irjet.net
3) Li, J., Li, B., & Lu, Y. (2021). Deep contextual video compression. In Advances in Neural Information Processing Systems (NeurIPS).
14) Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J., Cremers, D., Reid, I., Roth, S., Schindler, K., & Leal-Taixé, L. (2020). MOT20: A benchmark for multi object tracking in crowded scenes. arXiv. https://doi.org/10.48550/arXiv.2003.09003
4) Mentzer, F., Toderici, G., Minnen, D., Hwang, S. J., Caelles, S., Lucic, M., & Agustsson, E. (2022). VCT: A video compression transformer. In Advances in Neural Information Processing Systems (NeurIPS).
15) Kim, Y., Jeong, H., Yu, J., Kim, Y., Lee, J., & Jeong, S. Y. (2023). End-to-end learnable multi-scale feature compression for VCM. IEEE Transactions on Circuits and Systems for Video Technology. https://doi.org/10.1109/TCSVT.2023.3348527
5) Cheng, Z., Sun, H., Takeuchi, M., & Katto, J. (2020). Learned image compression with discretized Gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 7939– 7948). https://doi.org/10.1109/CVPR42600.2020.00796
16) Enabling next-generation consumer experience with feature coding for machines. (2025). In Proceedings of the IEEE International Conference on Consumer Electronics (ICCE). https://doi.org/10.1109/ICCE63647.2025.10930026
6) Yoon, Y.-U., Kim, D.-H., Lee, J., Oh, B. T., & Kim, J.-G. (2023). An efficient multi-scale feature compression with QP-adaptive feature channel truncation for video coding for machines. IEEE Access, 11, 89401–89415. https://doi.org/10.1109/ACCESS.2023.3307404
17) Lu, G., Cai, C., Zhang, X., Chen, L., Gao, Z., Zhang, D., & Ouyang, W. (2020). Content adaptive and error propagation aware deep video compression. In European Conference on Computer Vision (ECCV).
7) Zhang, Q., Mei, J., Guan, T., Sun, Z., Zhang, Z., & Yu, L. (2024). Recent advances in video coding for machines standard and technologies. ZTE Communications, 22(1), 62–76.
18) Ma, S., Gao, J., Wang, R., Chang, J., Mao, Q., Huang, Z., & Jia, C. (2023). Overview of intelligent video coding: From model-based to learning-based approaches. Visual Intelligence, 1(15). https://doi.org/10.1007/s44267-023-00016-9
8) Eimon, M. E. H., Adzic, V., Kalva, H., & Furht, B. (2025). Emerging standards for machine-to-machine video coding. arXiv. https://doi.org/10.48550/arXiv.2512.10230
BIOGRAPHIES
9) Gadot, U., Shocher, A., Mannor, S., Chechik, G., & Hallak, A. (2025). RL-RC-DoT: A block-level RL agent for taskaware video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 12533–12542).
Dr. Samir Khanji Informatics and Communications Department, Faculty of Engineering, Ebla Private University, Idlib, Syria.
10) Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., & Wang, X. (2022). ByteTrack: Multi-object tracking by associating every detection box. In European Conference on Computer Vision (ECCV) (pp. 1–21). https://doi.org/10.1007/978-3-031-20047-2_1 11) Jocher, G., Chaurasia, A., & Qiu, J. (2023). Ultralytics YOLOv8 (Version 8.0.0) Computer software. https://github.com/ultralytics/ultralytics 12) Ding, K., Ma, K., Wang, S., & Simoncelli, E. P. (2022). Image quality assessment: Unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5), 2567–2581. https://doi.org/10.1109/TPAMI.2020.3045810 13) Dendorfer, P., Ošep, A., Milan, A., Schindler, K., Cremers, D., Reid, I., Roth, S., & Leal-Taixé, L. (2021). MOTChallenge: A benchmark for single-camera multiple target tracking. International Journal of Computer Vision, 129, 845–881. https://doi.org/10.1007/s11263-020-01393-0
© 2026, IRJET
|
Impact Factor value: 8.226
|
ISO 9001:2008 Certified Journal
|
Page 353