VCANet: Vanishing-Point-Guided Context-Aware Network for Small Road Object Detection

Advanced deep learning technology has made great progress in generic object detection of autonomous driving, yet it is still challenging to detect small road hazards in a long distance owing to lack of large-scale small-object datasets and dedicated methods. This work addresses the challenge from two aspects. Firstly, a self-collected long-distance road object dataset (TJ-LDRO) is introduced, which consists of 109,337 images and is the largest dataset so far for the small road object detection research. Secondly, a vanishing-point-guided context-aware network (VCANet) is proposed, which utilizes the vanishing point prediction block and the context-aware center detection block to obtain semantic information. The multi-scale feature fusion pipeline and the upsampling block in VCANet are introduced to enhance the region of interest (ROI) feature. The experimental results with TJ-LDRO dataset show that the proposed method achieves better performance than the representative generic object detection methods. This work fills a critical capability gap in small road hazards detection for high-speed autonomous vehicles.