Driving assistance system based on deep learning and traditional vision

Abstract

Relevant technologies such as computer vision and artificial intelligence are cheaper and easier to implement than detection technologies implemented by hardware such as lidar and radar. Cars are equipped with advanced intelligent driving assistance systems to prevent or reduce traffic accidents. In this context, this paper will identify and analyze the most important traffic lights, vehicles, and lane lines in traffic. Based on ImageNet pre-training, SqueezeNet builds fine-tuned network recognition traffic lights. Aims to achieve an assisted driving system that integrates deep learning and traditional vision. The final model size is only 7.84MB, the recognition accuracy is as high as 94.95%, and the processing speed is 12.4ms / frame. The single-frame processing speed of recognizer of YOLO v3 trained vehicle and classifier of B-CNN trained vehicle is up to 24.47ms. Using computer vision and mathematical operations, image perspective transformation, and polynomial fitting to analyze lane lines has the advantage of reducing cost.