Traffic Sign Detection using Feature Fusion and Contextual Information

Traffic sign detection based on image and video data is critical, which captures real-time traffic road information for autonomous vehicle. With the rapid development of CNN, more and more CNN-based detectors have promoted general object detection. However, these mainstream detectors still suffer from small object detection task because of small size and fuzzy representation. Traffic signs are representative small object on road scenes causing a rigid challenge for autonomous driving perception system. In this paper, traffic sign detection (TSD) is regard as a small object detection task. We propose a feature fusion method via cross-connection to enhance feature representation. In addition, contextual information searched by dilated convolution is also used to support small traffic sign detection. We have implemented our modules into Faster R-CNN and evaluated effectiveness of proposed method on Tsinghua-Tencent 100K dataset. Our experimental results prove that the feature fusion method via cross connection and contextual information improve detection result of small traffic sign.