Research On High Robust Multi-Pedestrian Tracking Algorithm Based On DeepSort Framework

Multi-pedestrian tracking is a hot topic at present,which belongs to the intermediate processing of visual perception.The main purpose of multi-pedestrian tracking based on vision is to continuously track multiple pedestrian targets,keep the ID of each same pedestrian unchanged,and predict the future frame state of each pedestrian.In this paper,aiming at the common scenarios more pedestrian tracking algorithm to study the problem of insufficient robustness,study more practical DeepSort more pedestrian tracking algorithm is presented,and its solution coupling for the three modules,state prediction module of target,the target feature extraction module,the module object matching,this paper respectively from the apparent target modeling and target state prediction module two aspects to improve the DeepSort algorithm.The main contributions are as follows:In order to solve the problem that the original DeepSort algorithm is not robust enough under the intersection,a multi-pedestrian tracking algorithm based on multi-feature fusion DeepSort algorithm is proposed.In the modeling stage of target representation,HOG feature and CNN feature of target pedestrian were extracted to establish the classifier model of target and background,and the image confidence map of target pedestrian to be tracked was obtained by classification,and the confidence map was put into DeepSort tracking algorithm for multi-pedestrian tracking.The performance of the improved algorithm was measured by using the MOT16 dataset,a special dataset in the MOT Challenge competition.The experiment proved that the proposed improved algorithm was more robust than the original algorithm.In order to solve the problem that the proposed multi-pedestrian tracking algorithm based on multi-feature fusion DeepSort is not robust enough in the nonlinear case,and the complex road condition at the intersection is prone to nonlinear,a multi-pedestrian tracking algorithm based on multi-feature and UKF fusion DeepSort algorithm is proposed.Due to the need of target tracking task pedestrians to forecast the future frame motion state,the kalman filter method can only be used in the original algorithm for relatively simple linear state of pedestrian environment target forecast,so it will be extended kalman filter algorithm for no trace kalman filter algorithm,and applied it in the improved tracking algorithm,using the scenario are MOT 16 data sets to performance evaluation.