Human-inspired Crossroad Recognition based on Key Local Regions

Crossroad recognition is an important part of unmanned navigation, which provides crossroad location information. Most existing works treat scene recognition as a classification problem and achieve good results. However, it is difficult for them to cope with scene changes, such as occlusions and seasonal changes. In this paper, we propose a human-inspired crossroad recognition method that detects some key local regions of the scene to recognize crossroads. The key local regions contain the unique features of the intersection, which are extracted by visualizing the features of the pre-trained crossroad classification network. Based on the detected key local regions, an inference model calculates the score of crossroad classification and output the recognition result. The proposed method can effectively recognize the crossroad under occlusion and other changing scenes. We collected the crossroad recognization dataset on our campus, and evaluate the performance of our method on this dataset. Extensive evaluation shows that our method achieves higher accuracy and satisfies the real-time requirement.