Trajectory Prediction Based On Improved Hidden Markov Model

With the wide use of video traffic monitor,vehicle detection,vehicle tracking,vehicle positioning,trajectory analysis and other technologies based on computer vision have become research hotspots.These technologies have broad application prospects in the field of transportation,such as traffic flow parameter extraction,vehicle trajectory prediction,intersection safety risk assessment,etc.Vehicle detection is the basic method to extract traffic parameters from traffic videos.The technical difficulty is to reduce the influence of occlusion in traffic videos on the accuracy of vehicle detection.Therefore,prior information of Kalman filter is introduced to carry out vehicle fusion detection and tracking.For large-scale traffic scenes,the distributed video network is adopted,and the consistency information fusion algorithm is used to reduce the detection error of a single camera,so as to extract more accurate vehicle trajectories.Aiming at the application of vehicle trajectory prediction in traffic,the conventional Hidden Markov Model is improved to make it suitable for trajectory prediction.The main work and innovations of this paper are as follows.1)Vehicle tracking fusing the prior information of Kalman filter under occlusion conditions.By using the prior information of Kalman filter and the new vehicle region description method,the vehicle region of the image is precisely morphologically processed to reduce the occlusion.The occlusion problem is divided into the occlusion between vehicles and the occlusion between vehicles and obstacles on the roadside.The prior information of Kalman filter is introduced to segment and merge in two kinds of occlusion problems,which improves the accuracy of vehicle detection.2)Accurate vehicle positioning based on information weighted consistency.In the distributed video network,information weighted consistency algorithm is used to solve the problem of detection error of a single camera.After several iterations,the distributed state estimation results are close to the centralized state estimation results,and the detection accuracy is improved.Further,the algorithm does not rely on the information processing center,and the computing needs are dispersed to each sensor,which improves the stability of the system and reduces the cost of setting up the video network.3)Video-based vehicle trajectory prediction by combining an improved Hidden Markov Model with Kalman filter.The structure of Hidden Markov Model is improved to make it more suitable for trajectory prediction and the corresponding learning algorithm and prediction algorithm are deduced.Kalman filter is used to predict the short-term trajectory,and the improved Hidden Markov Model is used to predict the long-term trajectory.Finally combine the two so that the prediction algorithm can perform well in both the short-term and long-term trajectory prediction.