高速公路车道级视频检测区自动设定方法OACSTPCD

The paper proposed an automatic configuration method for lane-level detection zones on ex-pressways based on video,addressing the issue of the inability of Pan-Tilt-Zoom cameras to automatically set up video detection zones,which affected the accuracy of traffic event recognition.Firstly,by ana-lyzing the features of U-Net and MobileNet series models,combined with structures such as depthwise separable convolution and inverted residuals,a lightweight and efficient R-Net series model was de-signed specifically for semantic segmentation of lane lines and drivable areas.On this basis,according to the characteristics of expressways'specific scene recognition tasks,an algorithm for lane lines and detection zones labeling based on connected component analysis was proposed,which achieved auto-matic configuration of lane-level detection zones.At the same time,in order to improve the accuracy of the labeling algorithm,two preprocessing methods,threshold processing and frame stacking,were introduced for the first time.Then,a quadratic equation was used to fit the lane labeling results,achiev-ing complete and smooth segmentation of the lane lines.The experimental results showed that the per-formance metric MIoU of the R-Net series models was close to traditional models such as SegNet and U-Net,but significantly reduced the number of model parameters and inner product operations.Among them,the segmentation performance metric MIoU of the R-NetV2 model reached 90.6%,which was only 0.4%lower than U-Net,but its model parameter count was reduced by 38.7%and inner product operation was reduced by 62.5%.Labeling the preprocessed semantic segmentation results resulted in an increase in lane labeling accuracy from 80.47%to 95.58%compared to traditional methods.