Human Detection,Tracking And System Realization Research:Evidence From Metric Learning

Object detection and tracking are two major tasks in the field of computer vision,which are widely used in robots,intelligent transportation,smart city and monitoring fields.In this paper,we focus on the problems and shortcomings of object detection and tracking.We discuss the difficulties and challenges in this field.We implement a visual perception system integrating human detection and tracking algorithm.The main contributions of this work are summarized as follows:1.A high-speed object tracking algorithm based on convolution regression network is improved.The basic neural networks are designed and compared with different structures.The model's generalization ability is enhanced through the analysis of motion model and data augmentation.We test our tracking algorithm on OTB public datasets and compare with other algorithms.Our proposed algorithm has high accuracy and real-time performance.2.A human tracking algorithm based on metric learning is proposed.A multi-scale features reorganization method is designed to improve the accuracy of human tracking at different scales.By introducing the triplet loss function in metric learning,we solved the tracking difficulties of background clutter and similar objects interference.Our proposed algorithm is robust through testing each tracking difficulty on the OTB public datasets.3.A mini real-time intelligent perception system is developed,which have the function of detecting and tracking human.Firstly,we propose an improved human detection algorithm based on SSD,which raises the recall rate of small objects by designing the prior box with kmeans clustering.Secondly,the detection algorithm and the tracking algorithm are fused by feature sharing and sequence training.Finally,we constructed the multi-tasking logic framework by compressing and accelerating the model.