Regression detection based adaptive inverse perspective mapping for vehicle arc lane detection

Abstract

Inverse perspective mapping (IPM) plays an important role in intelligent vehicles and machine vision, which can eliminate perspective distortion and get a top view for lane detection and location. However, on the curved slope, the changing road surface brings the changing characteristics of pitch Angle and camera height, which distorts the bending Angle of the lane, resulting in the decrease of IPM recognition accuracy. In order to solve this problem, an adaptive IPM method based on surface and vanishing point (VP) is proposed to eliminate perspective distortion of complex pavement such as flat pavement and curved slope. In the adaptive IPM method, VP is used as a feature reflecting road changes and detected by Residual Neural Network with Split-attention (ResNeSt-50). The adaptive IPM method uses the differential judgment mechanism of VP variation characteristics to automatically determine the road section and select the corresponding IPM model in real time. Simplified IPM is adopted in the flat section, robust IPM is adopted in the transition section, and full csipm is adopted in the circular slope section. Finally, the dynamic IPM experiment verifies the effectiveness of the adaptive IPM method in the whole driving process. The method can maintain the effect of removing distortion under different slope angles and reduce the complexity of lane detection.