Research On Multiple Object Tracking Method Based On Sequential And Hierarchical Features

Multiple object tracking is a hot-point on computer vision researches.A good feature expression which is suitable for multiple object tracking determines the accuracy in tracking algorithm.Different from other handcraft features,features extracted from deep neural network hold the most intrinsic characteristics.By designing appropriate network structure and unique loss function,neural network can satisfy most of the feature extraction requirements in different computer vision tasks.The positions and sizes of objects in each frame are given by an object detector as priori knowledge in multiple object tracking.Tracking algorithm connects the same objects and distinguishes others.Besides,the targets in multiple object tracking task are usually in the same category.Therefore,traditional features used on multi-classification do not meet the needs of multiple object tracking.This paper focuses on the feature extraction problem in multiple object tracking.Inspired by the self-learning concept in deep learning,an online feature extraction scheme is proposed in this paper,based on a conditional random field(CRF).The CRF model is transformed into a certain number of stacked auto-encoders with a new loss function.Features obtained with our method contain both continuous and distinguishable characteristics of targets.Besides,in order to solve the unstable detection problem,this paper uses two of the ensemble learning methods,random forests and gradient boosting,to mix the deep feature and video context spatio-temporal information together.Then a robust and fused feature which is satisfied with multiple object tracking algorithm is gotten.The parameter optimization and dimensional reduction strategy are researched.In the end,to solve the occlusion issues in multiple object tracking and enhance the tracking accuracy,this paper develops the min-cost-max-flow multiple tracking model,and applies the robust and fused feature in it.This model is solved by a dynamic programming algorithm to reduce the time complexity.we test and compare our method with others on MOT Benchmark,then explicit the advantages and problems to be improved.