Object Tracking And Activity Recognition In Surveillance Video

With the development of deep learning,intelligent analysis of surveillance videos has become a hot research direction in computer vision,involving object detection,multi-object tracking,and activity recognition etc.This thesis focuses on multi-object tracking,multi-object tracking of multi-camera and activity recognition in surveillance video.For multi-object tracking on VIRAT dataset,we implement a multi-object tracking method based on POI framework,improving it by the appearance,shape,and location similarities of objects.In order to extract the appearance features of the object,we propose an adversarial threshold network based on Gated-CNN,which uses the network to extract more discriminative re-identification features instead of deep features.The network has achieved excellent performance on Market 1501 and other datasets.For multi-camera multi-object tracking of vehicles,we propose a multi-object tracking framework for multi-camera vehicles.This framework adaptively clusters all trajectories of single camera multi-target tracking based on the connectivity graph to determine tracks and number of vehicles in the monitoring scene.The program has achieved good results in multi-camera multi-target tracking task of AICITY CHALLENGE 2019For activity recognition in surveillance videos,we implement a complete and feasible framework in the ActEV series evaluation.For ActEV-2018 evaluation,we use trajectories to spatially locate activities and sliding window for time positioning behaviors;For ActEV-PC evaluation,to improve the temporals liding window-based in ActEV-2018,we propose anchor-based time positioning method;Also we improve 3D ResNet-based video classification model in ActEV-2018,using a more lightweight P3D network.Excellent results in the ActEV series evaluation.