Real-time Detection and Recognition of Live Panoramic Traffic Signs Based on Deep Learning

With the hotness of unmanned driving and the challenge of traffic safety, traffic sign detection and recognition is the core issue of unmanned driving. It is of great significance to study it. At present, due to the application of deep learning, the object recognition method has achieved good results, but there are few researches on real-time recognition of small object in live panoramic high resolution images or videos. This paper proposes a method for small object detection and recognition in high resolution images, and applies it to the detection and recognition of traffic signs in panoramic pictures. Firstly, the data preprocessmg method for training dataset is proposed. Then a YOLOv3 algorithm based Traffic Signs Network(TSNet) is designed. After training and fine-tuning, the testing dataset with reasonable scaling is processed by the new sliding window algorithm proposed in this paper. The predictions for bounding boxes is processed by Delete Subsets- Nonmaximum Suppression(DS-NMS) to obtain the higher accuracy and real-time detection effect.