Research On Activity Detection Method Of Surveillance Video

With the advent of the Internet of everything era,how to realize intelligent analysis and processing of surveillance video has attracted more and more attention and become a hot research topic in computer vision.This thesis focuses on activity detection in surveillance video,involving object detection,multi-object tracking,activity classification,activity region detection,etc.This research topic is based on TRECVID ActEV activity detection evaluation task.In ActEV SDL 2020 evaluation task,an activity detection system is not only required real-time processing,but also needs to solve too large area problem of 3D activity detection which contains a large amount of background and missing detection in dense scenes.In view of the above requirements,this paper designed and implemented an activity region detection scheme in indoor scene.Firstly,a single-stage real-time tracking framework FairMOT is adopted.Then,based on the prior knowledge of offline data statistics and trajectory analysis,the pedestrian trajectory is extended to obtain the trajectory of activity candidate regions.An algorithm of training sample enhancement based on activity trajectory is proposed to solve the problem of insufficient training data and poor generalization of activity classification network of our system.The validity of the scheme is verified on the validation set of MEVA data set.In the ActEV 2020 evaluation,we improve the activity region detection method of our laboratory prophase.In view of the missing detections of 3D activity region detection method,soft-NMS is used to improve recall rate and mAP.Aiming at the problem of large scale variation of actors,a proportional balanced pyramid convolution is introduced to extract scale invariant features to improve the multi-scale fusion effect.To solve the problem that the detection frame is not accurate enough,different backbone networks are selected for different activities.Finally,the activity track connection algorithm is used to connect the activity candidate regions into activity tracks to provide activity spatial-temporal candidate regions for subsequent activity recognition.The improved activity region detection scheme is combined with the activity classification scheme of our team,and finally win the second place in the ActEV 2020 evaluation.Aiming at the real time requirement of ActEV SDL 2021 evaluation and the expensive calculation problems of the previous method,a more lightweight activity region detection scheme is designed and implemented.In the selection of backbone network,a more lightweight network structure is selected,and the recall rate and accuracy rate are improved through channel attention,spatio-temporal attention mechanism and metric learning.Finally,experiments are carried out on the validation set of MEVA data set to verify the effectiveness of our scheme.