A novel model based on deep learning for Pedestrian detection and Trajectory prediction

Abstract

Pedestrian trajectory prediction is faced with the difficulty of data set acquisition, and the traditional models considers a single pedestrian in isolation and ignores the influence of the target pedestrian's neighborhood. This paper presents a pedestrian prediction model that integrates pedestrian detection, multi-target tracking and circular neighborhood adding. Our method uses two different LSTMs to capture the interaction information between the target person and neighboring pedestrians. For the pedestrian interaction information, the circular neighborhood is used instead of the traditional rectangular neighborhood in the social scale. We evaluate the performance of four prediction models on three common datasets. The results reflect that the proposed method can solve the problem of manually labeling datasets, and the circular neighborhood can improve the accuracy of trajectory prediction.