| As a popular research direction in computer vision,pedestrian detection and tracking have wide application scenarios,especially in traffic flow statistics,driverless,virtual reality,human-machine interaction and public safety.In real world,the accuracy of pedestrian detection and tracking algorithms is directly affected by pedestrian’s non-rigid deformation,occlusion and light changes.With the rapid development of big data and deep neural network,pedestrian detection and tracking algorithm has made amazing progress.This paper mainly focuses on pedestrian detection and tracking,and proposes a pedestrian detection and tracking algorithm based on the combination of YOLO(You Only Look Once,YOLO)and RJMCMC(Reverse Jump Markov Chain Monte Carlo,RJMCMC).This paper adopts the algorithm framework of tracking by detection.Therefore,pedestrian detection is the cornerstone of pedestrian tracking,and its detection performance directly affects the quality of pedestrian tracking.Because pedestrian detection needs to consider the detection accuracy and detection speed,this paper proposes a pedestrian detection algorithm based on YOLO.On the basis of the Yolo algorithm,optimizing the aspect ratio of anchor box prior to solve the problem of too-small pedestrian caused by pedestrian movement characteristics and distance can effectively improve the situation of missing pedestrian detection by YOLO algorithm.The experimental results show that the mAP of the optimized YOLO detection algorithm,in the test set of MOT Challenge,is increased by 3%on average,and the improvement effect is significant.Generally speaking,Bayesian theory could solve the problem of pedestrian tracking,that is to estimate the state of pedestrian according to the observation information.This paper first reproduces the Kalman filter tracking algorithm.Because it is only applicable to linear Gaussian system models,and pedestrian movement is irregular and belongs to nonlinear Gaussian system models,the tracking effect is not ideal.The particle filter can handle the nonlinear tracking problem,so a particle filter based pedestrian tracking algorithm is proposed in this paper.The particle degradation and diversity loss of the basic particle filter algorithm limit the algorithm.MCMC can effectively make up the defects above through producing a group of particles with no right value.However,particle filter based on MCMC can not solve the problem of multi-space pedestrian tracking,since that pedestrians will randomly enter the scene to change the number of tracking,resulting in changes in pedestrian state variable space.RJMCMC constructs a Markov chain with variable state dimension to sample particles and track pedestrians.According to the characteristics of pedestrian movement,it defines six reversible jumping strategies including adding,deleting,staying,leaving,updating and interacting.The pedestrian tracking is realized by combining the observation information provided by the detector.Experimental results show that MOTA on the MOT15 test set is 32.1%,MOTP is 70.9%,MOTA on the MOT16 test set is 44.2%,and MOTP is 74.2%.Compared with the methods in recent years,this method has a good performance in tracking accuracy and precision. |
