International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 09 Issue: 04 | Apr 2022
p-ISSN: 2395-0072
www.irjet.net
Real-Time Pertinent Maneuver Recognition for Surveillance Srinidhi S1, Balasubramanian M2, Singamala Monisha3 , Yuvarani P4 1Student,
Dept. of Computer Science and Engineering, S. A. Engineering College, Tamil Nadu, India Professor, Dept. of Computer Science and Engineering, S.A. Engineering College, Tamil Nadu, India 3Student, Dept. of Computer Science and Engineering, S. A. Engineering College, Tamil Nadu, India 4 Student, Dept. of Computer Science and Engineering, S. A. Engineering College, Tamil Nadu, India ---------------------------------------------------------------------***--------------------------------------------------------------------2Associate
Abstract – Live video streaming is continuously produced
level schematic for a picture, several challenges have to be overcome. First, visual data must be analyzed and transformed into a format that represents the visual content effectively. It assumes that the camera is stationary. The example focuses on detecting objects. The rest of this paper presents a survey of “state-of-the-art” frameworks and their limitations and describes our proposed technique.
across industries including media, reconnaissance, marketing, and much more. Live events play an important role up to the minute in espionage. Contemporary advances in machine learning techniques have shown great interest in producing possible information for real-time events to deliver advanced user information or timely notifications. Pertinent maneuver recognition for surveillance uses or owns a 3D model network, ResNet-34, Kinetics 400 dataset, and uses YOLOv4 deep neural network techniques for discernment of the venture with optimal speed and accuracy. The ResNet-34 will work with still pictures and conjointly works with the live video stream. YOLOv4 is a real-time state-of-the-art consuetude object detection modus operandi. The kinetics dataset is a highquality, huge dataset for automated human maneuver recognition in videos. The object detection dataset consists of custom-trained images of the armaments that are presumed to be possessed by a person. Our action symbol is the least complicated in structure and provides accurate results and thus is utilized in CCTV footage to descry whether or not a person is possessing accouterments on the qui vive.
Data can be collected through numerous resources like sensors, accelerometers, still images or video frames. When collecting data through sensors, people are required to wear more than one sensor in their body parts that are locomotive. The raw data needs to be processed in different methods. The raw data is segmented and various features are extracted. This process can be a challenging task in the case of sensors. The deep neural networks used in this paper help to extract more important features and then are subjected to classification algorithms. The classification model is based on the custom-trained dataset and is used to identify actions and detect weapons. Hidden Markov Models (HMM), support vector machine classifiers, and feed-forward neural networks are some of the classification algorithms. Newfangled stratagem comprehends Recurrent Neural Networks (RNNs), Convolutional neural networks (CNNs), and Multi-Layer Perceptron (MLP). These approaches are more suitable and convenient as the processing time of the video is reduced.
Key Words: 3D model network, ResNet-34, Kinetics 400, YOLOv4, deep neural network.
1. INTRODUCTION Real-time pertinent maneuver recognition for surveillance is a progressive approach to the discovery of digital content and visualization makes it increasingly challenging to search, edit and access visual information. Research to enhance the representation and understanding of visual content, Content Based Image Retrieval (CBIR), has continued for decades. Generally, CBIR is named after two styles of visual elements: global and indigenous or native features. Global feature-based algorithms aim to see concepts in visual content as a full. Local features are alternative and have fewer benefits than global features. Local feature algorithms focus on key points and vivid picture ports that contain rich local information in the image. This is automatically detected using various icons, e.g., Harris corner and Difference of Gaussian (DOG). The Scale Invariant Feature Transform (SIFT) may be a promising lowlevel visual descriptor, which is invariant to scaling, translation, and rotation and likewise partially invariant to illumination changes and affine projections. To get the high-
© 2022, IRJET
|
Impact Factor value: 7.529
In this paper, an automated pertinent maneuver recognition system is developed, with object detection using a custom-trained dataset is developed. The system consists of both activity and weapon detection models. The activity recognition system is developed through the ResNet-34 algorithm with a kinetics dataset. The weapon detection system is developed using YOLOv4 (You Only Look Once), a custom-trained object detection model. Therefore, the main advantage of using Resnet model is that the problem of vanishing gradient is hammered out with less training errors comparatively.
2. LITERATURE SURVEY There are divergent researches on object detection and human maneuver recognition that exhibit sundry facets of this model. Wearable has the capacity to transform and modifies people’s life better. In this technology, they absorb and collect all data from users and their surroundings.
|
ISO 9001:2008 Certified Journal
|
Page 3565