Research on Adaptive Segmentation Algorithm for Highway Roadside Video Images Based on Unet Networks

Aiming at the existing highway roadside video analysis applications that can't adapt to the rotation of the ballbar and the difficulty that the monitoring area needs to be re-calibrated after the rotation, the adaptive segmentation algorithm is studied. We collect 2370 images of various highway line types, make fine labeled motorway, non-motorized lane, central isolation zone, road environment and other highway video image area segmentation samples, train Deeplabv3, Unet, Pspnet three kinds of deep learning semantic segmentation network respectively, and combine the test results with the comprehensive analysis to select Unet semantic segmentation network, the semantic segmentation prediction and the real intersection and merge. The ratio of semantic segmentation prediction and real intersection reaches 92.45%, detection accuracy reaches 96.32%, and recall rate reaches 95.77%. For different preset positions of the ball machine rotation, adaptive video monitoring to provide different regional coordinates, and according to different regions to allocate different detection capabilities, effectively solve the video analysis algorithms due to the ball machine rotation failure, as well as the need to repeatedly calibrate the problem of manual calibration, and significantly reduce the highway algorithms in different scenes repeated access to the video streams caused by the pressure of network transmission.