Research On Pedestrian Accurate Identification And Tracking Algorithm Based On Multi-feature Fusion

Target tracking plays a vital role in the application of computer vision,such as drone tracking,unmanned driving,motion analysis,marine rescue and video detection.In this paper,computer vision and machine learning methods are used to deeply study the tracking and recognition of pedestrians occluded for a long time.It is a very challenging task to complete the tracking and recognition of pedestrians occluded for a long time in a complex environment,and discriminative representation is a necessary condition to complete this task.Therefore,this paper deeply analyzes the changes of various features in different target tracking scenarios,and through experimental verification,it is concluded that the method combining depth features,color features and pose features as representations can obtain better results in the evaluation of public data sets.Subsequently,in order to adapt the representation to different tracking scenarios,this paper proposes an adaptive representation network(ARN)to dynamically combine color features,depth features,and pose features to design target representations.Then use Euclidean distance and locality-constrained linear coding(LLC)as the metric to complete the queue matching.In addition,this paper proposes a posture supervision module(PSM)to improve the accuracy of the target frame and introduce an IOU filtering module(IFM)to eliminate the detection frame redundancy.Finally,a dynamic representation tracker(DRT)combining ARN,PSM and IFM is designed.Considering the lack of long-term occlusion pedestrian tracking dataset in the current academic world,this paper constructs a pedestrian occlusion tracking dataset(POTD)for the performance evaluation of pedestrian occlusion tracking algorithms.The algorithm and 15 tracking algorithms in this paper are tested on 7 sets of video sequences in the public OTB-2015 dataset and POTD dataset.Experimental results show that,compared with other algorithms,this algorithmachieves higher accuracy in the task of tracking and identifying pedestrians who are occluded for a long time.