Activity Detection And Recognition In Surveillance Video

With the development of economy and science and technology,people’s demand for security become more and more intense.Intelligent security has become the rigid deman d of social life.Intelligent video surveillance technology has been widely concerned and become a hot research direction in the field of computer vision,involving object detection,activity recognition,activity detection and other related tasks.This thesis focuses on the detection and recognition of vehicle turning and pedestrian activity in surveillance video.For vehicle turning activity detection,in TRECVID ActEV-PC 2019 evaluation,a dense sliding window and spatiotemporal IoU matching method is designed to make full use of the spatiotemporal information of video to improve the detection effect based on the trajectory sequence obtained from vehicle detection and tracking,solving the problem of poor robustness based on trajectory analysis and key area scheme,and win the first place in the evaluation.The research work of pedestrian activity detection is based on TRECVID ActEV 2019 evaluation and ActEV-SDL 2020 evaluation.In ActEV 2019 evaluation,we focus on solving the problem of unbalanced samples of different categories.The focal loss function is used to improve the extended model of two-stage object detection Faster-RCNN based on 3D convolution,which improves the recall of detection and gets the third place in the evaluation.ActEV-SDL 2020 evaluation requires real-time performance,that is,the running time of the system can not exceed the video playing time.Therefore,a lightweight activity detection and recognition scheme for surveillance video is designed.The single-stage target detection model RetinaNet is extended to a three-dimensional model,thus,the area of activity is directly detected,which effectively improves the speed and effect of detection,and complets the evaluation.