On Highway Guardrail Detection Algorithm Based on Mask RCNN in Complex Environments

Abstract

Highway guardrail survey is an important task of highway management. In order to improve the accuracy of guardrail detection in complex background, Mask RCNN was introduced in combination with image preprocessing algorithm, Resnet101 was used as the backbone network, feature pyramid network (FPN) structure was used for feature extraction, and regional proposal network (RPN) was used to generate regional proposals for each feature map. The mask image of the guardrails was generated by Mask RCNN to realize the segmentation and detection of the guardrails. The average precision of guardrail detection of 200 test images was 94.38%, and the average recall rate was 93.8%, and the MIoU rate for instance segmentation was 85.83%. The experimental outcomes indicate that the guardrail detection algorithm based on the Mask RCNN can accurately segment and detect guardrails under complex environmental backgrounds, and has good universality and robustness.