Study On Multi-object Tracking In Intelligent Traffic Scene

Object tracking algorithm is one of the hotspots in the field of compute r vision research.It is a key technology in application scenarios such as video behavior analysis,scene understanding,traffic management and security prevention and control.It has fields in smart retail,intelligent security monitoring and driverless cars.A wide range of applications.As the country began to vigorously develop and build smart cities and intelligent transportation,the data volume of public transportation surveillance video has been greatly improved.How to effectively analyze and use the video big data of these traffic scenes to extract valuable information,thereby assisting in optimizing urban traffic and improving public safety.In recent years,many research scholars at home and abroad have analyzed and discussed the object tracking technology and proposed a series of research algorithms.However,due to the complexity of the intelligent traffic scene,the pedestrian has the characteristics of rigid and flexible rigid body,and the appearance is susceptible to wearing,scale,occlusi on,posture and perspective.The existing multi-person tracking algorithm is still difficult to apply in practical scenarios,so the research on pedestrian tracking is still a very challenging research topic.In this paper,the current frontier multi object tracking algorithm is studied.The basic problem of multi-target tracking algorithm and the current multi object tracking algorithm based on deep learning are studied and analyzed.The solution to the multi object tracking algorithm is fully studied and borrowed.Under the tracking framework of detection,a multi object tracking algorithm based on deep learning that can be applied to complex traffic scenarios is designed.The main work of the thesis is as follows: The multi-target tracking algorithm is deeply researched and analyzed.For the characteristics of pedestrian appearance and size change in traffic scene,this paper designs a pedestrian-based feature extraction network model based on metric learning.Based on the convolutional neural network model,pedestrians can obtain more accurate appearance characteristics,which improves the tracking accuracy of the pedestrian tracking algorithm.The target motion features in the existing multi object tracking algorithm are mostly based on the filtering algorithm.This paper attempts to design a pedestrian motion feature prediction network based on the siamense neural network,which alleviates the uniform velocity using the Kalman filter algorithm to some extent.The problem caused by the motion model is that the prediction accuracy is low.A multi object tracking algorithm was designed for the apparent features and the fusion motion features,and was evaluated on the MOT16 dataset.Finally,a multi person tracking experiment under real and complex scenes is carried out for the actual application scenario.Finally,the algorithm designed in this paper uses the detection data provided by POI,and the tracking performance MOPA and MOTP reach 59.3% and 61.3%,respectively.