Design and Application of Autonomous Driving System based on Multi-sensor Fusion

Abstract:
A method of multi-sensor data fusion using deep learning is proposed for the autonomous driving system. First, the realtime road condition information is collected through the binocular camera mounted on the car model, and the road and the obstacles on the road are identified with the semantic segmentation of the deep neural network image. At the same time, supplemented with other information collected by the laser sensor and ultrasonic sensor on the car model, as well as the road information obtained by Beidou navigation, the Kalman filter parameters are optimized with deep learning and passed via algorithm fusion to the car model actuator for automatic operation. The signals collected by multiple groups of the same type of sensors on the car model are optimized by an improved Kalman filter algorithm to achieve fast and accurate output of real-time results. The actual vehicle experiment results of the smart car model show that the solution can realize tasks such as autonomous navigation and driving, automatic obstacle avoidance, automatic route planning and prediction in a variety of complex road environments, with an operating speed equivalent to 80 km/h and a safety rate of 99.8%.