Research On Pedestrian Tracking And Trajectory Prediction Methods In First-person Video

Today's society has gradually entered the "artificial intelligence" era.A large number of new technologies are driving the development of information and intelligence in the society,bringing many technologies including robotics,speech recognition,image understanding,natural language processing,and intelligent recommendation systems.At the same time,computer vision,as an important part of the artificial intelligence neighborhood,has become a hot research content.It researches computer vision-related theories and technologies,and then builds artificial vision systems that obtain "information" from image or video data.Make computers replace human eyes and brains to perceive,interpret and understand the real world.Pedestrian tracking and trajectory prediction,as hot research topics in the neighborhood of computer vision,have been successfully applied in areas such as intelligent security monitoring and blind navigation,and have achieved good development.However,unlike third-person pedestrian tracking and trajectory prediction,the first-person video has low tracking and trajectory prediction accuracy due to frequent occlusions,interactions,and scene changes.In order to improve the accuracy of pedestrian tracking and trajectory prediction,this paper focuses on the first Pedestrian tracking and trajectory prediction methods in one-person video to better meet the needs of practical applications.First,a pedestrian detection model with improved single-step multi-frame detector(SSD)is proposed to solve the problem of low pedestrian detection accuracy due to frequent pedestrian occlusion in first-person videos.This method first adjusts the network structure of the SSD to change the feature extraction strategy.Second,it resets the default box and uses a default box with an aspect ratio to make predictions.Finally,it uses the repulsive force loss function to calculate and obtain accurate pedestrians.The detection results better solve the problem of false detection and missed detection caused by frequent pedestrian occlusion.Then,in order to solve the problem that the tracking accuracy is not high due to the lack of pedestrian characteristics in the first-person video and the interaction between pedestrians,a pedestrian tracking method combining deep apparent features and social power optimization is proposed.This method uses the pedestrian detection method based on the improved single-step multi-frame detector to obtain accurate pedestrian detection results,extracts the deep apparent features of the detected images,uses the Kuhn-Munkres algorithm for data association,and finally optimizes the association results based on the social force model.To improve the accuracy of pedestrian tracking.Finally,a pedestrian trajectory prediction method combining attention and scene characteristics is proposed for the problems of pedestrian interaction and real-time transformation of moving scenes in first-person video.The encoder-decoder structure of LSTM is used to capture the influence of prior information by encoding its own motion information.It uses the attention to quantify the pedestrian impact and uses the scene features extracted by the CNN model to represent the scene impact.Trajectory prediction has higher prediction accuracy.