Research On Background-Aware Filter Based Integrated Tracking Algorithm

As an important branch of computer vision,object tracking has been widely used in various fields of real life,such as intelligent monitoring,human-computer interaction,and driverless vehicle.Although the object tracking field has accumulated a lot of excellent results by a long period of development,there is currently no algorithm that can adapt well to all tracking scenarios due to the complexity of tracking scenes,such as object deformation,scale change,occlusion,motion blur,etc.So there is still a difficult task how to improve the robustness of the tracking algorithm to complex scenes.In this paper,the current mainstream correlation filtering algorithms are studied.Based on the current mainstream correlation filtering tracking framework,this paper aims to improve the background-aware correlation filters algorithm for challenges of target tracking in complex scenarios from four aspects:feature extraction,scale prediction,occlusion processing and integration processing.The main work is as follows:First,this paper introduces feature fusion strategy and fast scale filter to improve the performance of background-aware correlation filters algorithm for the problem of single feature and inaccurate scale estimation.Firstly,a variety of complementary features are introduced into a strong feature in the feature extraction stage of the algorithm,and so the characterization of the feature to the target is enhanced.Secondly,the performance of the algorithm is improved by introducing a fast scale filter to predict the target scale separately.Second,a parallel tracking algorithm based on feature fusion and fast scale filtering is proposed for the occlusion problem.By analyzing the peak characteristics of the correlation response map in the normal tracking situation of the tracker and the occlusion of the object,a response map quality measure is proposed for occlusion determination.And a parallel tracking strategy is proposed to distinguish between object occlusion and large deformation.Third,an improved multi-cue integrated tracking algorithm is proposed for the problem that the single tracker is easy to drift and the feature utilization is low.The algorithm ensemble multiple base trackers with different features.The proposed tracking filters comprehensive robustness evaluation metric is used to evaluate tracking results,which can effectively determine the best tracking results.Through the feedback of the final result,the mutual error correction mechanism between multiple trackers is realized,and so the robustness of the algorithm is effectively enhanced.Also,a lot of experiments on the improved parallel tracking algorithm and multi-cue integrated tracking algorithm based on background-aware filter are carried out on the OTB-100 benchmark dataset and compared with other mainstream algorithms.Experiments show the effectiveness of the improved method proposed in this paper,and the improved algorithm has better tracking performance.