Perspective Distortion Model for Pedestrian Trajectory Prediction for Consumer Applications

Predicting human motion and interpreting the trajectory of a pedestrian is necessary for consumer electronics applications ranging from smart visual surveillance to visual assistance of autonomous vehicles. The majority of existing work in trajectory prediction from camera sensors as input has been investigated mostly in the top-down view (ETH and UCY datasets). However, accurate prediction of pedestrian trajectory used in first person/third person view of visual surveillance and autonomous driving is still a challenging task. With the increasing deployment of these IoT devices and the integration of AI for decision-making, human trajectory prediction can significantly contribute to improving consumer experiences and safety in these contexts. In this article, we propose a lightweight geometry-based Perspective Distortion Model (PDM) that leverages first-person/third-person view property of perspective distortion for long-term prediction. The qualitative result shows a promising prediction of future positions with 2, 3, 4, 6 seconds in advance over videos taken at 30 fps. Our proposed model quantitatively achieves state-of-the-art performance in terms of the Average Displacement Error (ADE) while tested on a self-created dataset (https://github.com/RahulRaman2/DATABASE) and Oxford Town Centre dataset.