Research On Multi-objects Tracking Algorithm Based On Deep Learning

Multi-Object Tracking(MOT)is one of the most popular research directions in the field of computer vision and has many important applications including intelligence surveillance,robot navigation and autonomous driving.It is also an important part of building smart city.In china,there are a large number of monitoring cameras,which create a lot of videos per second.How to implement efficient retrieval and object tracking in such large number of videos is a difficult problem.A stable and efficient object tracking system is necessary.Many researchers have proposed a series of tracking algorithm from different perspectives.However,due to the complexity of multi-target tracking,such as object occlusion and moving camera,current algorithms are difficult to have high robustness and high performance at the same time,and can not be applied in real scenes.In existing MOT methods,the tracked objects usually show consistent or slowly varying appearance across time.Visual features of the objects are therefore important cues for associating detection boxes into tracklets.In recent years,deep learning techniques have shown great potential in learning discriminative visual features for single-object and multi-object tracking.However,when tracked objects with similar appearances occlude or are close to each other,visual cues alone cannot guarantee robust tracking results.To tackle those difficulties,this thesis proposes a Deep Continuous Conditional Random Field(DCCRF)for solving the online MOT problem in a track-by-detection framework.The contribution of this online MOT framework is two-fold:(1)estimate tracked objects' displacements across time based on visual appearance information.They are modeled as deep Convolution Neural Networks.Instead of previous linear motion model or quadratic motion model,this algorithm can be applied to more scenes.(2)In existing MOT methods,individual objects' movements and inter-object relations are mostly modeled separately.In addition,inter-object relations are mostly modeled in a symmetric way,which this thesis argues that is not an optimal setting.High-confidence tracklets should be used to help correct errors of low-confidence tracklets and not to be affected by low-confidence ones much.The robustness of the proposed method is improved.Moreover,the DCCRF is trained in an end-to-end manner for better adapting the influences of visual displacement prediction based on neural networks as well as inter-object relations based on continuous conditional random fields.The difficulty of training is greatly reduced.Extensive experimental comparisons with state-of-the-arts on benchmarks demonstrate the effectiveness of this MOT framework.