Research On Target Tracking And Wandering Behaviour Detection Methods Based On Video Surveillance

With the increasing advances in computer vision technology,the processing of surveillance video data is becoming increasingly intelligent.Traditional surveillance processing methods to obtain the behavioural trajectory of a particular target require a lot of manual effort and time to search and filter through the vast amount of data to find valid information,which is no longer able to cope with the massive amount of video data required in today’s society.Intelligent surveillance systems,with their high efficiency and accuracy,are rapidly replacing traditional surveillance systems and are of great importance to the intelligent development of social security.In practical applications,it has become fundamental to safeguard personal safety and public security to quickly,accurately and directionally detect dynamic targets in surveillance scenes where abnormal behaviour occurs and to continuously track them to control potential risks.To address the above problems and realistic requirements,this thesis firstly adopts an improved YOLOv5 network for scene detection to obtain pedestrian targets,and uses a pedestrian tracking model based on the Deep SORT algorithm in continuous video frames to correctly correlate the detected targets for tracking,and finally proposes a wandering behaviour detection method based on the dynamic conversion scale factor of the object image on the basis of continuous tracking,for identify whether abnormal behaviour occurs in pedestrians and feeds back into tracking,so that numerous targets appearing in the scene can be tracked quickly and accurately in a focused manner.The main research elements of this thesis are as follows:(1)An improved solution to the YOLOv5 target detection algorithm is proposed.A Squeeze Net network compression and expansion method is used in the benchmark network YOLOv5,and the fire module is added to replace part of the convolutional block structure,thus acquiring more feature information while reducing the computational effort.The attention mechanism is also embedded into the CSP2＿X module in the Neck section,allowing the network to focus on information with a greater degree of relevance in the current task objective,thus improving the feature representation and accuracy of the network.(2)The key steps of pedestrian tracking and the construction method of the model are investigated.A pedestrian tracking model is constructed using the improved YOLOv5 network as a detector combined with the Deep SORT algorithm.The model uses the improved YOLOv5 network to implement the detection part of the pedestrian target,then predicts and updates the tracking frame based on Kalman filtering,and correlates the data between the detection frame and the tracking frame by the Hungarian algorithm to achieve pedestrian tracking in surveillance video.(3)A wandering behaviour detection method based on the dynamic conversion ratio coefficient is proposed.The motion characteristics of the target when wandering behaviour occurs are summarised and analysed,and the directional characteristics are defined.The coefficient of dynamic conversion is then introduced to determine the direction and speed of pedestrian movement according to its rate of change,and when the value of the rate of change fluctuates between positive and negative,the pedestrian is judged to have wandered,and when the value is zero,the pedestrian is judged to have stopped.Experimental validation of the above method and model was carried out on the INRIA detection dataset,MOT16 tracking dataset and pre-defined pedestrian movement patterns.The experimental results show that the accuracy,recall and average precision of the improved YOLOv5 network are improved by 1.74%,4.45% and 2.56% respectively,reducing the occurrence of missed and false detections and making pedestrian detection more effective.The designed pedestrian tracking model has a lower ID conversion rate,and the tracking accuracy and tracking precision are improved by 1.3% and 3.2% respectively compared to the traditional Deep SORT,which can meet the tracking requirements of intelligent surveillance.The wandering behaviour recognition results based on the dynamic conversion scale factor of the object image match the preset pedestrian movement pattern to detect wandering and stopping behaviour during pedestrian movement,providing a new method for the research and application of pedestrian movement behaviour discrimination.