Integrated light-ResNet and Poolformer networks for shape-preserving lane detection

The automatic Lane detector will be a useful safety-reliable apparatus in the autonomous vehicle if it can suppress the misses, fakes in lane detections to help unveiling the car-intruded lane-occlusions just in time. By the video captured in adverse environments, the road lane distortions including occlusion, rain reflections, heavy shadows, and paint-peeling can often let the lightweight lane detector fall in puzzling for a while, hence certainly impact the self-drive safety. Hence, in this study, the lane detection networks composed of a primary detector, which is the launched ultra-fast lane detector using RestNet16, and a prediction-refiner based on Poolformer is introduced for recovering the occluded portions and the natural smooth shape of curve lanes. The Poolformer refiner contains a single channel-MLP according to a coarse-to-fine recursive training scheme. When the difference of inferenced patterns between the primary detector and the Poolformer refiner exceed the tolerable risk threshold, the replacement is bypassed. Hence, the proposed refiner in our system such as is a revision-conditional rectifier to actually achieve appropriate and effective refinement. The simulation results demonstrate that the proposed rectifier in the revision-conditional way can rapidly recuperate the lane marks for adequate structure integrity to benefit the self-driving safety.