Height Restriction Bar Detection via Deformable YOLOv5s

In recent years, there are more and more traffic accidents caused by the height restriction bars, as current vehicles are becoming taller and taller. To avoid such traffic accidents, automatic height measurement system for height restriction bars is urgently required. To accurately measure the height, the location of the height restriction bar should be detected first. However, to the best of our knowledge, there are no professional algorithm and relevant dataset for height restriction bar detection. Therefore, in this work, we elaborately collect a novel dataset for height restriction bars, named as HRB22, which comprises more than 8600 images covering various types of height restriction bars in different natural scenes. This dataset could serve as a catalyst for promoting many vision tasks, e.g., segmentation and height estimation, etc. As we know, the height restriction bar is usually thin and has long boundary. Therefore, a deformable module is proposed based on the modulated deformable convolution to facilitate the learning of these features. In order to satisfy the real-time detection requirement for driving, the proposed feature learning module is integrated into the backbone of YOLOv5s. A variety of experiments and ablation studies are conducted to demonstrate the effectiveness of the proposed dataset and deformable module. In addition, the proposed method is tested on a common object detection dataset MS-COCO, which illustrates the generalization of the proposed method.