Pedestrian counting is a significant research field in computer vision. An important pedestrian counting technique is the detection-based approach, which counts the number of pedestrians under surveillance. However, the directions of pedestrian movement in surveillance videos are not uniform. Therefore, in traffic scenes and public places, pedestrian counting methods considering movement directions are particularly crucial. To address this issue, this paper proposes a trajectory-based pedestrian counting method for counting pedestrians moving in different directions. The approach first employs YOLOv5 for pedestrian detection, using the detection results as initial positions for tracking. DeepSort is then utilized for pedestrian tracking to obtain movement trajectories. Subsequently, we analyze the obtained trajectories. In this phase, we first calculate the initial direction of the trajectory to determine the pedestrian’s movement direction. Then, at each following frame, we detect whether a U-turn or turning behavior occurs. We model turning behaviors as inflection points. Finally, the count value for the current frame is adjusted based on the detection results. Our method is evaluated on the public pedestrian detection dataset. The average precision (AP) of our pedestrian detection method is 94.9%. The average accuracy of the pedestrian count is 93.2%. The experiments prove that the proposed method performs well in pedestrian counting according to the direction of movement.
Direction-based Pedestrian Counting Method Leveraging Trajectory Analysis
16.05.2025
16213592 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Europäisches Patentamt | 2024
|Pedestrian Counting Methods at Intersections
Transportation Research Record | 2007
|