Railway Traffic Signal Recognition System based on Spatio-Temporal Features

Automatic recognition of signals occupies an essential place in railway self-driving systems. Signal recognition systems help drivers avoid human factor accidents from fatigue and distraction of train drivers. In computer vision, developing object recognition algorithms is an important area of research related to transportation. This field is evolving exponentially with the help of deep learning, high-performance cameras, and reliable datasets. Nevertheless, the complexity of the states and combinations of railway signals has led traditional object recognition algorithms based on a single image as a dataset to add more structure to the recognition of one of the states of the signals described as "Blinking". This paper investigates a neural network based on YOLOv8 with a recurrent mechanism that extracts spatio-temporal contextual features expressed by scintillation signals. This research compensates for the lacunae in the railway transportation industry in terms of deep learning based approach in recognizing scintillation traffic signals. We enhance the training speed with a dual-channel parallel training structure and design an Ultra Fast Feature Extraction Module. Our network model implements real-time inference on 20fps video.