Detection of Rarely Occurring Behaviors Based on Human Trajectories and Their Associated Physical Parameters

The complexity of detecting rarely occurring behaviors through human trajectories is closely related to a lack of data, unclear behavioral characteristics, and complex variations in their related physical parameters (e.g., velocity and orientation angles, etc.). In this context, we propose a methodology to maximize the detection performance of rarely occurring behaviors in public places by investigating the data collection process, trajectory representation based on detected skeleton poses from videos, and the use of 2D (X, Y) trajectory positional data only versus its combination with their associated physical parameters as the input for trajectory learning models. In order to evaluate the proposed method, we studied a rare Japanese behavior in public places called UroKyoro, which is a combination of the two Japanese words Urouro and Kyorokyoro. This behavior includes aimlessly moving while frequently looking in both directions. Since there is a lack of related data from real-life cases, we hired professional actors to role-play the behavior alone or with normal pedestrians moving around. The learning system was trained using limited and augmented data. The trajectory learning system, trained with combined human trajectories and orientation angles following the proposed method, succeeds in detecting the studied behavior with an accuracy of 91.33%, outperforming the accuracy of the trained model using only human 2D (X, Y) trajectories by 4.33%. The results show the effectiveness of the proposed method to detect complex, rarely occurring human behaviors by training the LSTM classifier with a combination of human trajectories and physical parameters. However, the effectiveness of physical parameters on training performance may differ from one case study to another based on behavioral characteristics.