Traffic Signal Recognition Using End-to-End Deep Learning
Tonmoy Kumar Sarker, Xiangyu Meng · 2022
Traffic signal recognition is a core task in autonomous driving. In this paper, an end-to-end, real-time deep learning object detection algorithm called YOLO (you only look once) is adopted for traffic signal recognition applications. Two separate open-source image data sets were used for training signal detection models for day and night-time traffic scenarios, respectively. Since the default YOLO framework offers real-time detection, focus was given on achieving acceptable levels of detection accuracy through improving data quality for training. First, the training data sets were cleansed of existing ground truth annotation errors and irregularities. Second, by applying mosaic image augmentation, stitching four randomly upscaled images together, twice during the model training, overall detection accuracy was further improved.