Optimizing Self-Driving Car Navigation in India: Real- Time Traffic Police Gesture Recognition using MoveNet-Thunder

This study introduces a novel Traffic Police Hand Gesture Recognition System specifically tailored for autonomous vehicles operating in India, utilizing TensorFlow's MoveNet Thunder model. Due to the unique traffic management in Indian urban environments, where traffic police frequently use hand signals to direct traffic, autonomous vehicles face significant challenges in interpreting these signals correctly. This paper details the development of a deep learning-based system designed to recognize and interpret traffic police gestures such as 'Stop', 'Turn Left', 'Move Forward', and 'Turn Right'. The methodology involves the creation of a custom dataset comprising 8,000 labeled images of traffic police in diverse environmental conditions, which was used to train the MoveNet Thunder model. The system integrates camera feeds with LiDAR data using sensor fusion techniques to enhance recognition accuracy, especially in poor visibility conditions. Extensive testing in the Carla simulator demonstrated the system’s efficacy, with an achieved accuracy of 89% in gesture recognition. The results indicate a significant improvement in the interpretability and reliability of autonomous vehicle responses to manual traffic signals. This research contributes to safer and more efficient autonomous vehicle navigation by enabling accurate real-time interpretation of traffic police hand gestures, highlighting the potential for widespread application in similar urban settings globally.