Autonomous driving technology is profoundly reshaping the future travel mode, and environmental perception ability is a prerequisite for vehicle decision-making planning, and its reliability is directly related to driving safety. Compared with high-cost multi-lidar schemes, camera-based computer vision technology has become a key technology for the large-scale popularization of automatic driving due to its rich scene semantic information and hardware cost advantages. Based on this, this paper discusses how to build a vision-dominant perception module with high robustness, high efficiency and high safety. It is expected to provide technical reference for the landing of low-cost and high-performance automatic driving perception schemes.