Vehicle pressure line detection based on improved Mask R-CNN + LaneNet

To address the problem of vehicle pressure line detection of on-board images in the field of assisted or automatic driving, as well as the problem of missed and false detection caused by underexposure, shadow, or solid occlusion in the detection process, a vehicle pressure line detection algorithm based on improved Mask R-CNN and LaneNet was proposed. In terms of network optimization, based on the Mask R-CNN network, the image scaling algorithm (bilinear interpolation) of the ROI alignment layer was improved to bicubic interpolation, and the convoluted VGG16 network of a full connection layer was replaced by LaneNet's E-Net shared decoder. For image enhancement, the Gamma correction algorithm was improved to realize the automatic correction of underexposed images. In terms of training data, the vehicle target in the Tusimple data set was marked, and the data were enhanced in the network training process, based on the improved random erasing algorithm. The experimental results show that while the vehicle detection speed remains unchanged, the lane line detection speed is increased by 28%, and the vehicle missed and false detection rate are reduced by 38.93% and 89.04%, respectively. Further, the lane line missed and false detection rate are reduced by 67.21% and 87.05%, respectively. The achieved performance index can meet the requirements of vehicle line pressing judgement method. © 2022, Science Press. All right reserved.