Research On Extraction Of Pedestrian Track Information And Abnormal Behavior Detection And Analysis Based On Video

Public security issues are becoming increasingly prominent as a result of social development and technological advancement.Video surveillance,as a main means of real-time access to public area information,how to locate abnormal events quickly and accurately based on video surveillance has become a key problem to be solved.This thesis takes the video captured by the surveillance camera as the input and the pedestrian as the subject to be monitored,and carries out the research according to the idea of pedestrian detection and tracking,pedestrian trajectory information acquisition,trajectory model design,abnormal behavior detection and analysis.The following are the main work of this thesis:(1)An improved trajectory information extraction algorithm based on Yolo X and Deep-SORT is proposed.In the pedestrian detection stage,a strategy of dynamic data enhancement and optimization of the bounding box regression function is proposed to improve the accuracy of the pedestrian detection model in order to address the problems of low recall rate and unstable positioning accuracy of the target detection model.Deep-SORT is used as the foundation for multi-target real-time tracking.Based on this algorithm research,it has been discovered that when pedestrians are occluded,it causes trajectory breakage and mismatching.As a result,the calculation of trajectory confidence is introduced in this thesis as the basis for determining whether pedestrians are blocked.Experiments show that the proposed algorithm can improve the phenomenon of target switching and frame loss,and solve the problems of trajectory fracture and false detection effectively.The algorithm achieved 83.2% tracking accuracy on MOT16 dataset,and the detection accuracy of pedestrian detection model on VOC2007 dataset reached 86.13%.Experiments show the effectiveness of the proposed algorithm in target detection and tracking,which provides a strong guarantee for subsequent research.(2)A model for detecting abnormal behavior based on random forests and parameter optimization is created.First,the Douglas-Peucker algorithm is used to compress the problem of redundant trajectory original data.Then,representative trajectory features(discrete curvature entropy,displacement,distance,main direction Angle,abnormal growth of trajectory coordinates in x and Y directions and velocity changes in X and Y directions)are designed according to the performance characteristics of abnormal behaviors.Finally,the design features are verified with weak correlation and non-redundant,and an anomaly detection model based on random forest algorithm is established.Cross-validation is used to optimize model parameters,and the best results are used as model parameters to improve model detection performance.The CASIA behavior analysis database is re-divided in this thesis,and the model validity is compared and verified based on the dataset.Experiments show that the detection accuracy of the abnormal behavior detection model designed in this thesis reaches 96.7% on the above dataset,and 94.01% on the UCF-CRIME dataset.