A Fast Skeleton-Based Recognition of Traffic Police Gestures with Spatial-Temporal Graph Convolutional Network

Abstract

In recent years, driver assistance systems have been improved to assist the driver. One of the essential and indispensable features is traffic police action recognition. It would be applied in the field of self-driving cars - one of the areas of interest to the research community and is being developed widely in several major countries around the world, or the driver assistance system. This study proposes a fast recognition of traffic police actions based on human skeleton characteristics. Our proposed model will first detect the joints of the human body using the MediaPipe algorithm, then feed them to a Spatial-Temporal Graph Convolutional Network (ST-GCN) in order to classify the police actions into eight basic categories: Stop, Move Straight, Left Turn, Left Turn Waiting, Right Turn, Lane Changing, Slow Down, and Pullover. The experiments conducted on the real Traffic Police Gesture Dataset have shown the effectiveness of our proposed method.