Steering Angle Estimation for Self-driving Car Based on Enhanced Semantic Segmentation

Common approaches for semantic segmentation using Convolutional Neural Networks (CNN) based around conventional U-shapes architectures were widely used. However, failures to retrieve global context information and memory issues made such models unable to compete against modern architectures considering accuracy and real-time capability. In this paper, an efficient method maintaining equivalent accuracy of a previous segmentation network and skillfully making a model more light-weighted for real-time inference was proposed. More concretely, we managed to alleviate five out of 17 million trainable parameters, which effectively reduce the amount of computation of the original PSPNet by 30% using the backbone of CSPNet. Our proposed network implementation achieved 73 mIoU scores on our custom dataset and reached 15 fps regarding real-time inference. We deployed the trained model on the multifunctional hardware and then connected it to a golf car to jointly navigate the natural environment and traffic sign detection task. Accordingly, the STM32 board and servo motor were used for controlling the steering wheel through a track-and-wheel drive system. As for the traffic sign detection task, we employed a small-size Yolov5 trained on the TT100K dataset running around 60fps and attained real-time performance with sufficient accuracy.