International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 11 | Nov 2025
p-ISSN: 2395-0072
www.irjet.net
“Real time object detection” Prof. Shobha S Biradar1, Sunil2 1
Professor, Master of Computer Application, VTU, Kalaburagi , Karnataka ,India Student , Master of Computer Application ,VTU, Kalaburagi , Karnataka ,India -------------------------------------------------------------------------***---------------------------------------------------------------------efficiency. Furthermore, real-world environments often ABSTRACT- Real-time object detection is a rapidly evolving 2
field within Artificial Intelligence (AI) and Machine Learning (ML) that focuses on the automatic identification and localization of objects in images or video streams with minimal delay. Unlike traditional image classification, which only determines the presence of an object, object detection provides both classification and spatial information using bounding boxes. Recent advancements in deep learning, particularly Convolution Neural Networks(CNNs) and architectures such as YOLO (You Only Look Once), SSD (Single Shot Multibox Detector), and Faster R-CNN, have significantly improved the accuracy and speed of detection.
present additional difficulties such as variations in lighting, object occlusion, background clutter, and camera noise, which make detection tasks more complex. Therefore, the problem addressed in this project is the development of a reliable real-time object detection system that leverages advanced AI and ML techniques, particularly deep learning models, to overcome the shortcomings of traditional approaches and provide accurate, fast, and efficient object detection for real-world applications.
Keywords: Real time object detection, Deep learning, Real-time object detection, Computer vision, object detection.
The main objective of this study is to design and implement a real-time object detection system that can accurately identify and classify multiple objects in live video streams using Artificial Intelligence (AI) and Machine Learning (ML) techniques. To achieve this, the study is guided by the following specific objectives:
3. OBJECTIVES
1. INTRODUCTION This unified detection approach frames the task as a single network that directly regresses object bounding boxes and class probabilities from full images, enabling extremely fast end-to-end inference suitable for real-time video. It trades some localization finesse for speed but set the standard for practical, deployable detectors on live streams. Because the model is compact and trains end-to-end, it became a popular baseline for edge and camera-side applications where throughput is critical. Its simplicity also made export and engineering tooling easier, accelerating adoption. Use this family as the first production candidate when latency and simplicity are top priorities.
4. METHODOLOGY USED The methodology of this study outlines the systematic approach followed to design and implement the real-time object detection system 1) Model Selection: Choose a suitable pre-trained deep learning model such as YOLOv8 or SSD for real-time performance. 2) System Design: Develop a pipeline where video frames are captured from a webcam or video source.
This single-shot detector predicts object boxes and classes from multi-scale feature maps, handling different object sizes in one pass and offering an attractive speed/accuracy tradeoff. It demonstrated that proposal-free detectors can rival two-stage systems while remaining faster—useful when you need more accuracy than the smallest real-time models but still require high FPS. The multi-scale default-box mechanism is widely reused in mobile and embedded detectors. For mid-tier hardware, SSD is a pragmatic option balancing precision and throughput.
3) Implementation: Integrate the chosen model into the system using Python and OpenCV. 4) Testing and Evaluation: Test the system under different conditions such as lighting variations, moving objects, and occlusions.
5. LITERATURE SURVEY Mobile / lightweight detectors (various, 2016–2023) — Mobile Net-SSD, Tiny-YOLO, and Efficient Det-lite variants show how architecture choices (depth wise separable convs, smaller feature maps) minimize compute and memory while preserving usable accuracy. These models are tailored for CPU/mobile/NPU inference where full-size detectors are impractical. Use them for on-device, battery-sensitive applications.
2. PROBLEM STATEMENT The primary challenge lies in designing a system that can accurately detect and classify multiple objects in a video stream while maintaining real-time performance. Achieving this requires balancing speed, accuracy, and computational
© 2025, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 754