A Method for Re-identifying Subjects in Video Surveillance using Deep Neural Network Fusion

• Published in: International Conference on Computing, Communication, and Networking Technologies (Online) pp. 1 - 4
• Main Authors: Wanaskar, Ujwala H., Dangore, Monika, Raut, Dipak, Shirbhate, Radha, Borate, Vishal Kisan, Mali, Yogesh Kisan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24.06.2024
• Subjects: Cameras; convolutional neural networks; deep residual networks; Feature extraction; Machine learning algorithms; mask rcnn; Object recognition; person re-identification; Real-time systems; Residual neural networks; Security; Streaming media; Video surveillance; Visualization
• ISSN: 2473-7674
• DOI: 10.1109/ICCCNT61001.2024.10726255

Video security has become increasingly important in the contemporary world, primarily in response to a rising number of undesirable incidents. Video surveillance enhances security measures in various sectors of society. Identifying and re-identifying individuals in motion, particularly those observed across multiple cameras, presents a substantial challenge, further complicated by factors like videos with limited visual detail, fluctuating lighting conditions, and densely populated environments. Digital image processing elements including a model databases, feature descriptors, and classifiers are necessary for video-based individual identification. In particular, real-time object identification for human detection is the goal of many machine learning techniques. This work aims to present a new method for person re-identification (Re-ID) from video footage and an algorithm used for the task. Person Re-ID research employs diverse methods recognizing person by features, using multiple components matching for detailed profiles, and employing multiple component dissimilarity to capture appearance nuances. An impressive technique combines Mask RCNN and deep residual networks (DRN), known for their exceptional effectiveness in extracting appearance-based features. Following component extraction, these elements can be converted to empower the correlation and matching of people's appearances across changed video casings and camera points. The paper presents an inventive profound learning-based approach for individual Re-ID. It starts by distinguishing individual in a video outline utilizing Cover RC-NN and therefore extricates individual appearance highlights with D RN. The amalgamation of these two networks is used to reidentify person. The paper also addresses key challenges in person Re-ID in video surveillance and suggests potential solution.