Multi-layer LSTM Traffic Police Gesture Recognition Integrating Limb Angle Features and Attention Mechanism

Abstract

According to international self-driving technology standards, if self-driving vehicles are to be driven on the road, they must have the function of recognizing traffic police gestures. At present, traffic police gesture recognition methods are mainly divided into three categories, namely recognition based on bioelectric signals, sensor-based recognition, and machine vision-based recognition. This thesis mainly focuses on the situation that traditional machine vision technology easily ignores key coordinates and temporal features when processing dynamic traffic police gestures. This thesis proposes a multi-layer LSTM model that integrates the continuous sub-limb angle and attention models of traffic police. Based on Mediapipe, after unifying key points, the model trained with fusion of angle information has a higher accuracy than the model trained without fusion of angle information, and the model trained with 33 key points and their angle information of Mediapipe is more accurate than 501 key points and their angle information. Finally, based on the model proposed in this thesis, good test results were achieved on the Chinese traffic police gesture data set.