Multiagent Sensor Integration and Knowledge Distillation System for Real-Time Autonomous Vehicle Navigation

Abstract

This article introduces a comprehensive multiagent prototype system designed to enhance the autonomous navigation capabilities of vehicles by incorporating numerous sensors and components. The system includes features such as an ultrasonic sensor for precise distance measurement, a specially crafted "SonarSpinner" with a wide 160 degrees field of view, a vision sensor for road sign detection and steering angle estimation, and an infrared obstacle avoidance sensor, operating with a predefined obstacle-halting threshold of 150 cm. Data collection for model training and evaluation is accomplished using a virtual reality-based self-driving car simulator, resulting in a diverse dataset. The proposed system harnesses knowledge distillation from teacher models, such as the Nvidia model, to create a lightweight student model optimized for real-time inference while retaining competitive accuracy. Additionally, a custom Haar cascade classifier enhances traffic sign detection capabilities. The distilled model is then converted to TensorFlow Lite for efficient deployment on edge devices within autonomous vehicles, ensuring a secure and efficient navigation system. This innovative approach combines optimized distillation methods with specialized classifiers to facilitate the development of robust and real-time self-driving car systems.