A Study On Multi-object Tracking In Road Scenes

Multi-object tracking is an important research topic in Compute Vision.In recent years,single object tracking methods based on correlation filter has attracted people's extensive attention due to their fast speed and excellent performance.In addition,the feature extraction method based on deep learning has been widely used for object detection and tracking with its' high precision.The multi-object tracking algorithm based on Markov Decision Process is able to combine detection,single object tracking and data association technology effectively,and take advantage of each module.Therefore,correlation filter-based single object tracking algorithms and deep learning theory are applied in the multi-object tracking method based on Markov Decision Process in this thesis.Detailed works are as follow:First,in the 'Tracked' state of Markov decision process,the single-object tracking module is implemented by the classical correlation filter method KCF(Kernelized Correlation Filter)with a state transition policy based on high confidence detection,in which the advantages of KCF are exploited to improve the speed of multi-object tracking.Second,in the 'Lost' state of Markov decision process,recurrent neural network and Hungary algorithm are utilized in the data association module.By careful design of the feature extraction network and feature fusion of detections and tracklets,the problems of high computational complexity and slow running speed of neural network are improved.Finally,based on different appearances of pedestrians with different interval durations,a time-division data association technique is proposed.The performance of data association module is further improved,which facilitates data re-association in multi-object tracking.In summary,based on the multi-object tracking framework by Markov decision process,The speed of tracking algorithm and the performance of multi-object tracking are improved by the use of fast correlation filter single object tracking algorithm and high-performance neural network technology.Multi-object tracking accuracy is increased by nearly 5%compared to the baseline method.The motheds and results in this thesis can be used for intelligent surveillance,assisted driving and so on.It has important significance for the subsequent research of multi-object tracking tasks.