Row Classification Based Lane Detection for UAV Images

This paper specifically focuses on lane detection from the UAV's perspective. In comparison to traditional vehicle-based lane detection, the background information in UAV images tends to be more complex, offering relatively less road information but more interference. Moreover, lane targets in UAV scenes are generally narrow and elongated, resulting in higher computational costs for detection. Considering these unique characteristics, we frame the task of lane detection as a row-based classification problem, where we classify the anchors in each row of the image into lane or non-lane categories, and present a lane detection network based on row classification. Our approach utilizes the ResNet-34 as the backbone and incorporates the CBAM. Additionally, during the network training phase, we incorporate an auxiliary segmentation network to augment the network's capacity in capturing visual features of a lane. To validate the efficiency of our network, we perform ablation study and conduct comprehensive comparisons using a self-made UAV aerial lane dataset. The results of the ablation study and the comparison experiment affirm that our method successfully achieves a balance between accuracy and real-time performance in lane detection, making it a feasible solution for lane detection tasks involving UAV images.