Paper Title
Temporally Aware Move Net Thunder – Real Time Recognition of Traffic Police Gestures in Indian Autonomous Navigation Systems

Abstract
This Traffic Police Hand Gesture Recognition System for autonomous vehicles builds on the TensorFlow’s MoveNet Thunder model. The hand gestures recognized in the system include 'Stop', 'Turn Left' and 'Move Forward', all common gestures used by traffic police in India. The custom dataset consisted of 8,000 images with different gestures under diverse conditions. We employ dense and dropout layers in our neural network architecture to achieve a high accuracy and avoid overfitting. Haar cascades are used for real time face detection, so gestures are recorded only when the officer is facing the camera. The accuracy of the model is 89%, while in most cases of the classes it is well across, but there were a few minor misclassifications between similar gestures. The system was validated using the Carla simulator with and without weather and lighting conditions. A prototype of this solution is shown to be promising for safe and efficient integration into autonomous vehicle systems for navigation in traffic-managed environments. Keywords - Traffic gesture recognition, Autonomous vehicles, MoveNet Thunder, TensorFlow, Real-time detection, Carla simulator, Sensor fusion, Indian traffic