Video surveillance: Item monitoring against theft

Chong, Yiing Sheah (2023) Video surveillance: Item monitoring against theft. Final Year Project, UTAR.

Preview

PDF Download (61Mb) | Preview

Abstract

This research aims to address the limitations of traditional Closed-Circuit Televisions (CCTVs), which lack innovative and intelligent computer vision and video analytics capabilities that can protect human assets. Besides, the conventional one would rely on human operators to be constantly observing the scenes captured so that crimes could be discovered earlier and therefore prevented. Here, we have developed an intelligent video surveillance system that is capable of identifying the close loitering event and monitoring the select item in the scene by integrating these approaches, such as the You-Only-Look-Once (YOLO) version 5 object detection algorithm, close-contact detection, feature matching, significant movement and dissimilarity detections. Firstly, the system will preprocess the incoming frame if necessary to reduce the computational resources. The initial frame of the video will be saved as a background frame for tracking purpose. Then, the system operator or owner is required to manually select the item being monitored in the scene by drawing a bounding box for isolation purpose. The selected object will be utilized for initial feature extraction and descriptor computation. Next, the system will invoke the person detection module using YOLO every 1 second to detect the presence of human. If there is human being detected in the scene, the system will check whether the human is having a close-contact with the registered item by finding the intersection area of the human’s bounding box and the registered object’s bounding box. If the two objects are found intersected, the close contact time will be added by 1. Else, the time will be set to 0. After that, the system will get the current frame to find the common features with the background frame by using a feature matching algorithm. The respective counter for matching will be increased by 1 if the average match calculated is under certain thresholds. Then, the system continues to track the matched feature points in the current frame by calculating the optical flow. If significant changes in the positions of the matched feature points have been detected, the motion counter will be incremented by 1. After that, the system identifies the dissimilarity between the background and current frames to detect the presence of occlusion and the absence of the registered item. If the dissimilarity is greater than the threshold defined, vii the respective counter will be increased by 1. All the mentioned counters will be set to 0 if the conditions specified are not met. Based on the predetermined conditions for each type of risk, the appropriate alarms will be activated while a risk occurs. These alarms serve as an early warning for the system operator or owner, allowing them to take any necessary actions to address the risk. After testing, our developed system has demonstrated an impressive accuracy of 92.5% in detecting abnormal behaviours in 40 video inputs.

Actions (login required)