| With the rapid development of economy,science and technology,the demand of intelligent monitoring system is increasing,which is becoming an important measure to protect society and people’s livelihood.Therefore,the development of intelligent monitoring analysis technology has also become concerned.In the field of computer vision,the key technologies of activity analysis for surveillance videos involve many related vision tasks,such as object detection,activity classification,activity detection and so on.This thesis focuses on the classification of person-only activities and vehicle-only activities,and the detection of person-vehicle interaction activities in surveillance videos.In view of the complex activities and large number of categories in the evaluation task of TRECVID ActEV 2020,the existing classification models have low classification accuracy.Based on the activity trajectory sequence obtained by previous activity detection and tracking,the classification frameworks of human,vehicle and person-vehicle interaction activity are designed,which make full use of spatio-temporal domain information of videos to improve the classification effect.And the classification performance is also improved by solving the problem of sample imbalance of different classes.We won the second place in TRECVID ActEV 2020 evaluation.Based on TRECVID 2021 ACTEV-SRL evaluation task,the complexity of person-vehicle interaction activity in MEVA dataset makes it difficult to detect.We design and implement a person-vehicle interaction activity detection scheme.The X3D-Cascade-RCNN model improved from X3D and Cascade-RCNN networks is used for activity detection,which improves the model precision and recall rate.A two-stage personvehicle interaction classification module is designed to achieve the best performance of person-vehicle interaction detection.During the evaluation,the evaluation metric Pmiss@RFA=0.1 of the person-vehicle interaction is 0.5314,which ranks the first place. |
PDF Full Text Request