| Pedestrian object tracking is a widely studied issue in machine vision.It is widely used in video surveillance,human-computer interaction,and autonomous driving.However,the current research content is mainly for single object tracking,in the real environment,there are often many pedestrians.Multi-object tracking needs to face many problems,such as the dynamic changes in the number of targets,the occlusion of targets,the appearance of new targets and the disappearance of old targets,and the clutter of tracking scenes.It is still a challenging task to design an efficient and accurate pedestrian multi-object tracking algorithm.Pedestrian multi-object tracking is usually divided into four steps: pedestrian detection,feature extraction,similarity calculation and data association.The core steps are target detection and data correlation.That is,the pedestrian target is detected on each frame,and the result of the target detection is used to correlate to obtain the target tracklet.The data association depends on the extraction of features,such as appearance features,motion features,and time features.In order to improve the accuracy of pedestrian multi-object tracking,this thesis conducts research around the steps of feature extraction and data association,and proposes a pedestrian multi-object tracking method,which adopts the feature fusion extraction method based on boundary regression and combines with online data association method.This method simultaneously performs feature extraction and data association in a single CNN network,and uses tracklet confidence to manage pedestrian targets,which greatly improves the problem of tracklet interruption and tracklet transformation caused by pedestrian miss detection and pedestrian occlusion.Through the experiments in the MOT dataset,it is proved that the algorithm in this thesis has greatly improved the accuracy of pedestrian multi-object tracking and has reached the advanced level in the industry.The details are as follows:1.In order to solve the problem of missed detection and false detection of pedestrians,this thesis adds feature extraction branches to the classification and regression network based on feature pyramids and uses pedestrian boundary regression to predict motion trajectories.So that the model can adapt to different scales of pedestrians and greatly reduces the impact of pedestrian misses and false detections on tracker performance.2.In order to avoid the problem of target track drift and false association caused by occlusion,this thesis proposes a data association method based on twin networks.This method uses non-maximum suppression to select the best candidate frame from pedestrian detection and motion prediction and uses the apparent characteristics of the best candidate frame to build a data association matrix for learning,thereby reducing false associations and improving the accuracy of multi-object tracking.3.In order to solve the problem of tracklet ID transformation caused by long-term occlusion,this thesis adopts the tracklet confidence method to form a reliable pedestrian tracklet by using feature weighting to update the associated trajectories.Based on the above improvements,this thesis builds a pedestrian multi-object tracking framework to achieve accurate and efficient pedestrian multi-object tracking in videos. |
