Research On Multi-pedestrian Motion Tracking

Multi-object tracking has attracted much attention in the field of computer vision since its important academic value and extraordinary commercial potential.For example,it has been widely applied in many fields such as autonomous driving,aerospace,medical diagnosis,and intelligent monitoring.In the past ten years,researchers have proposed plenty of innovative algorithms to address this subject.The technology of multi-object tracking has made great developments,but it still faces major challenges such as overlapping targets,dramatic changes in appearance,light changes,and identity switch.We aim to study the multiobject tracking from the perspective of data association.We use pedestrian re-recognition technology to fusion multi-hypothesis tracking and graph matching respectively.We have achieved some innovative research results in the way of constructing the hypothesis tree,solving the matching results,and solving the identity switch and occlusion problems of multi-object tracking.The main contributions are summarized in the following two aspects:1.Although multi-hypothesis tracking considers all possible connection relationships and obtains the best trajectory,it results in slow operation and high memory consumption,when it faces complex scenes(such as crowded scenes and occlusions),there is no historical information to assist and causes false associations and identity switch.In order to alleviate the above two problems,this paper proposes a multi-cue multi-hypothesis multi-object tracking algorithm based on re-identification which takes pedestrian re-identification feature discrimination,metric learning,and correlation filtering as clues to fusion,and builds a hypothesis tree for multi-hypothesis tracking to tap the potential of multi-hypothesis tracking.At the same time,the historical information is retained that is used to re-identification the lost target,which effectively alleviates the exponential growth problem caused by too many hypothesis branches in multi-hypothesis tracking,reduces the identity switch,solves tracking failure,and improves the accuracy of the multi-object tracking algorithm.2.The existing graph-based deep learning methods mostly enhance the feature discriminability of a single target,but ignore the relationship between the targets,which makes the model unable to mine the information between the targets.In complex scenes,it is difficult to distinguish between similar targets,and the number of objects tracked by multiple targets is uncertain,which makes it difficult to use neural networks to achieve the process of multiple object tracking.In order to solve these two problems,we design an end-to-end fully differentiable framework based on graph convolution for pedestrians.This is a new paradigm for dealing with data associations between trajectories and targets,and between targets and targets.We use depth graph matching to find the corresponding relationship between the nodes in order to maximize the similarity of the corresponding nodes and edges,that is,the similarity between the targets of the same identity is the largest,and the relationship between the targets and the targets are mined,which is effective improve the accuracy of matching.We adopt end-to-end training,and all modules of this method cooperate with each other.This method reduces the parameter adjustment work,thereby improving the model adaptability.