Pedestrian Direction Estimation: An Approach via Perspective Distortion Patterns

Abstract

Knowledge of pedestrian's walking direction is very crucial in multiple domains of video processing. This paper proposes a graph based, robust and light weighted model for direction estimation of pedestrian's walk by using the property of perspective distortion. Here perspective distortion pattern is used as an advantage in estimation of direction. The graph-based solution uses 3 parallel approaches for estimating the direction: Perspective Distortion Graph, Centroid Displacement and Clustering of Vanishing point. A pedestrian in a frame can be identified by bounding boxes. The temporal dimensional features of bounding boxes are height and width and these features changes for a particular object from frame to frame as the objects moves. These changes are unique for each direction for each object. These changes in dimension along with clustering of vanishing point and centroid displacement is used for the assesment of the pedestrian's walk direction. All the existing approaches need some sort of pre-processing on the frames, which makes the model more complex and time consuming. In the proposed model, the video sequence is applied on YOLO V4 algorithm and bounding boxes are obtained. By analysing the changes from frame to frame for the dimensions, graphs are plotted and minimum and maximum extremas are detected form the graph by eliminating soft extremas. After that envelope is placed for the graph and an average line is drawn based on the envelope, which will give the inference about the direction of walk of the pedestrian. The perspective distortion graph will not give accurate estimation for all directions. So, Centroid displacement and clustering of vanishing point are also used for direction estimation. The result obtained from the three methods are combined and form a robust model. For accurately estimating walk direction, the movement is limited to 8 different directions. For experiment, NITR Conscious Walk dataset and self-acquired dataset are used. With balanced accuracy of 97.003% and 96.25% and a false positive rate of 0.63% and 0.65%, respectively, the model produces good results for the above dataset.