Font Size: a A A

Surveillance Video Object Detection And Activity Detection

Posted on:2021-12-07Degree:MasterType:Thesis
Country:ChinaCandidate:Y LiFull Text:PDF
GTID:2518306308968449Subject:Electronics and Communications Engineering
Abstract/Summary:PDF Full Text Request
With the advancement of science and technology in recent years,the construction of smart cities and safe cities has been promoted throughout the country.The intelligent processing of surveillance video plays an important role in the construction of smart cities.Activity detection is the core of intelligent video processing system,and object detection is the basis of activity detection.This article will focus on object detection and activity detection in surveillance video.In terms of object detection for surveillance video,this paper uses Faster-RCNN as the basic model,analyzes the characteristics of surveillance video data,and makes improvements in feature selection,bounding box regression,and difficult sample discovery based on the difficulties.The improved model has been improved on the ActEV-DET object detection dataset constructed from the ActEV evaluation dataset.The activity detection in surveillance video is the focus of this article.The research work of activity detection is based on ActEV-2018 evaluation and ActEV-PC evaluation.In the ActEV-2018 evaluation,the research focuses on the detection of human-vehicle interaction.This paper proposes a trajectory-based two-stage human-vehicle interaction detection model.The model first proposes candidate behavior trajectories based on rules,then slices and classifies candidate trajectories,and finally post-processes the behavior start and end times.In ActEV-PC evaluation,in view of the complexity and diversity of activities,this paper proposes a three-stage activity detection framework,which can efficiently and effetively detect various behaviors.The three phases of the detection framework are spatial localization,activity area connection,and temporal detection.During the spatial localization phase,Faster-RCNN was extended to a three-dimensional model based on 3D convolution,and the activity area was directly detected,which effectively improved the spatial localization effect.The detection framework achieved first place in the ActEV-PC evaluation.
Keywords/Search Tags:Activity Detection, Object Detection, Deep Learning
PDF Full Text Request
Related items