Low light recognition of traffic police gestures based on lightweight extraction of skeleton features

Abstract

With the increasing demand for information interaction between intelligent vehicles and traffic managers, understanding traffic police commands has become an essential element in achieving human–computer interaction. The aim of this study is to investigate the low light traffic police gesture recognition based on lightweight extraction of skeleton features. A two-stage recognition framework for eight gesture commands is proposed. The first stage utilizes convolutional neural networks to acquire the human skeleton under low light. The second stage utilizes a recurrent neural network to construct a lightweight feature extraction model for gesture recognition. HRnet acquires spatial features of the action, high-resolution and multi-scale fusion circumvents the effect of low-light intensity. Self-attention (SA) mechanism and MobileNet lightweight feature extraction have addressed the time-consuming issue in the feature extraction phase. The improved GRU network is used to extract time domain features, the fusion of spatio-temporal features improves the recognition accuracy. For training and validation, a gesture dataset (TPGD) containing 3391 gesture instances is constructed. The proposed method obtains 94.75% accuracy in the TPGD test set, demonstrating its effectiveness. This study holds significant importance for enhancing the level of intelligence in traffic management, promoting the development of intelligent transportation systems, and ensuring road traffic safety.