Pedestrian Trajectory Prediction And People Tracking From Multiple Cameras

In recent years,the number of video surveillance equipment in public places such as airports,subways and shopping malls has grown rapidly.However,checking abundant video data with human vision has great difficulties.It is of great academic and commercial value to use computers,utilizing some intelligent methods,to process,analyze and mine information from videos.In this paper,we focus on the topic of pedestrian trajectory prediction and multi-camera multi-pedestrian tracking in computer vision.The goal of multiple cameras multiple people tracking is to determine the position of each person at all times.Pedestrian trajectory prediction is to predict the future trajectory of pedestrians based on the historical observation.They can be applied to video surveillance and analysis,motion and crowd behavior analysis,autonomous driving or suspicious and abnormal situation detection.However,research related to multiple cameras and multiple targets tracking has many challenges,such as long-term occlusion,great changes in perspective,camera setting and pedestrian's posture,which makes it depend more on reliable appearance features.When people walk in crowded spaces such as sidewalks,subways and airports,they naturally adjust their walking style according to the context of the scene and follow common social etiquette,such as maintaining isolation and avoiding collisions.These implicit interactions will lead to complex group movements and bring about great challenges to pedestrian trajectory prediction.In pedestrian trajectory prediction,we propose a data-driven model based on graph neural network,which simultaneously infers the interaction between pedestrians in an unsupervised manner and predicts the trajectories of all pedestrians at the same time in a crowded scene.We propose a new multi-camera multi-people tracking method and a new pedestrian trajectory prediction method.In multi-cameras and multi-people tracking,we use a two-branch deep neural network to extract target's appearance features.We fuses global appearance feature and body semantic map to get more robust appearance feature.Based on the feature,we assemble a multi-cameras multi-people tracking processing pipeline.Through extensive experiments on public datasets of multi-cameras and multipeople tracking and pedestrian trajectory prediction,we conduct performance evaluation of the implemented method.The results show that our multi-cameras and multi-people tracking model is superior to existing tracking methods.Compared with existing trajectory prediction methods,our trajectory prediction model gets better performance in terms of tracking average displacement error and final displacement error.In sum,this research provides new ideas and solutions for problems related to multi-people tracking,and also provides support for higher-level application systems.