Abnormal Event Detection in Video Surveillance Using Yolov3

Abstract

The importance of the automated detection of abnormal occurrences within video streams has grown in tandem with the proliferation of video surveillance equipment. When compared to the typical course of events, an abnormal occurrence may be thought of as a departure from the norm. Despite this, the ratio of normal to abnormal occurrences is highly unbalanced due to the fact that abnormal occurrences do not happen very often. A technique for detecting video abnormal events that is based on CNN (convolutional Neural Networks) as well as instance based learning has been suggested. This approach was developed in detection to the need that video abnormal events be localized in pixel-level regions. First, the Gaussian background models are used to precisely pinpoint the moving targets inside the movie, and then, using an image processing approach, the associated areas of the pinpointed moving objects are acquired. Finally, the pre-trained is put to use in order to extract features from linked areas, which are then utilized to generate multiple kernel learning packages. In the end, the multiple instance learning model learns using the normalized set kernel approach, and pixel-level predictions are made. Deep learning and the You Only Look Once v3 (YOLOv3) object detection technique are going to be combined in this model for the detection of highway accidents. According to the findings of the experiments, the technique of video anomaly detection that is based on CNN and sparse representation learning is able to precisely discover abnormal occurrences in the area that is comprised of pixels. Using this as a use case, the purpose of this thesis work was to provide an initial solution for the same problem using deep learning techniques. This was done to avoid the need to involve human resources in the monitoring of any abnormal activities that were spotted in the live broadcasts from the monitoring system