Research On Pedestrian Motion Trajectory Prediction Method Based On Visual Tracking

With the rapid development of computer vision technology,intelligent applications such as automatic driving,video surveillance and robotic autonomous navigation rely more and more on the results of accurate and effective pedestrian motion analysis.Pedestrian motion analysis and pedestrian behavior understanding have gradually become research hotspots.Especially in the field of intelligent transportation,capturing pedestrians within a certain range and prejudgement of their motion are of great significance for improving road traffic safety.However,as a kind of vulnerable road uers in the traffic scene,pedestrians' irregularities,the complexity of the background,and the uncertainty of movement all make it difficult for the analysis of pedestrian motion and pre-judgment of pedestrian status.Even after decades of development,computer vision-based pedestrian motion analysis technology has made some breakthroughs,but it still leaves much room to improve.Therefore,the study of pedestrian trajectory prediction based on visual tracking has a high theoretical and practical significance.This thesis studies the pedestrian trajectory prediction method based on visual tracking.The main contents are as follows:(1)Pedestrian detection method: Improved Aggregate Channel Features(ACF)pedestrian detection method based on improved AdaBoost and Non-Maximum Suppression(NMS)is proposed.Based on the ACF pedestrian detection algorithm,the classical AdaBoost algorithm in ACF is replaced by the robust alternating AdaBoost algorithm.For the NMS algorithm in the original ACF algorithm,variables such as size quotient,detection score quotient,and Outer Window Retention(OMR)are proposed.The OMR mechanism is used to fuse candidate detection windows to form a new ACF algorithm,and the improved ACF is used to train classifiers and detect pedestrian in images.The experiments on the public dataset and self-acquired pedestrian image dataset,it is verified that the improved ACF has better robustness and reduces the pedestrian miss detection rate and false detection rate.(2)Pedestrian tracking method: Improved pedestrian tracking method based on improved Tracking-Learning-Detection(TLD)and state transition policies is proposed.For the problem that the optical flow method(Lucas-Kanade,LK)in the TLD tracking module is easy to drift during the tracking process,an extended LK method based on a simple template updating mechanism and occlusion processing logic is introduced to repeatedly extend LK on multiple image scales.The method enhances the convergence range of the objective function and forms an improved TLD algorithm.After detecting the pedestrians in the image using the pedestrian detection algorithm proposed in(1),the improved TLD algorithm is used for pedestrian tracking,and the life cycle of each pedestrian is performed by the Markov Decision Process(MDP)Modeling,finally,the Hungarian algorithm is used to correlate the tracking results with the detection results to obtain the motion trajectory formed by the pedestrians in the process.Experiments on self-collected pedestrian video datasets on public datasets demonstrate that the improved multiple object tracking algorithm has good results.(3)Pedestrian trajectory prediction method: Pedestrian trajectory prediction method based on Long Short-Term Memory(LSTM)and visual attention mechanism is proposed.On the basis of the LSTM network,the social-distance aware pooling layer is added to obtain the SD LSTM,and the interaction between pedestrians in the local neighborhood is modeled.Then,the time and space map is used to introduce a visual attention mechanism to globally change the movement state of all pedestrians in the crowd.The model predicts the trajectory of pedestrian motion based on the aforementioned pedestrian tracking.The experiments on the public dataset and the selfacquired pedestrian video dataset,relatively small trajectory prediction errors are obtained,and the pedestrian trajectory prediction is realized to some extent.This thesis improves the accuracy of pedestrian detection by improving the ACF algorithm.By improving the tracking module in the TLD and using the improved TLD for multi-target tracking based on MDP,the tracking accuracy is improved.Finally,the pedestrian historical location information is input into the trajectory prediction network based on the social LSTM and attention mechanism,and the predicted pedestrian position is obtained,which improves the prediction accuracy.In this thesis,pedestrian detection,tracking,and motion trajectory prediction are combined into a set of sequential processes,and a pedestrian motion analysis technique is explored.