Real-Time Recognition Method of Vehicle Taillight Signal Based on Video Understanding

Vehicle taillight signals contain a wealth of semantic information crucial for inferring the leading vehicle's driving intentions. In this paper, to enhance the recognition accuracy, optimize the hardware requirements for deployment, and reduce the model reasoning time, a lightweight method for taillight signal recognition is proposed. This method involves three stages: detection, tracking, and recognition. Firstly, the lightweight MCA-YOLOv5 network is designed for the detection of vehicle rears. Subsequently, the detection results are tracked via the Bytetrack algorithm, resulting in tracking sequences. Finally, the TSA-X3d network aims to effectively obtain spatio-temporal information from the tracking sequences and recognize the taillight signals. The experimental results indicate that the MCA-YOLOv5s network significantly outperforms its precursor—the original YOLOv5s model—in efficiency and size. Specifically, the model size, parameter count, and computational demand of the proposed MCA-YOLOv5s network are respectively reduced to 33.33%, 31.43%, and 34.38% of those of the original YOLOv5s, yet it maintains comparable average precision. Furthermore, when compared with other typical taillight signal recognition algorithms, the TSA-X3d network not only has the fewest number of parameters, but also achieves the highest accuracy, reaching 95.39%. To mitigate deployment challenges, the study leverages TensorRT to markedly decrease the inference time of the MCA-YOLOv5s model to 1–3 ms. Additionally, it employs quantization techniques on the TSA-X3d model, slashing its inference time to 30% of the original and shrinking its model size to 26.65% of its initial footprint. Notably, the augmented algorithm's inference speed surpasses 25 frames per second, surpassing the threshold necessary for real-time taillight signal recognition, thus showcasing its potential for immediate practical application.