Vibration Visualization: High-Quality Vehicle Detection With DVS Data Using Deep Learning Network

Perception of vehicle operating status plays a crucial role in improving traffic flow efficiency and safety. However, due to the influence of complex environmental factors and traffic conditions, the efficiency and progress of traffic flow monitoring have been difficult to improve. To address this issue, this paper proposes an end-to-end vehicle vibration recognition method based on distributed optical fiber vibration sensing technology (DVS). Firstly, a DVS system for real-time vehicle monitoring is established to obtain vehicle vibration information. Subsequently, the acquired DVS data is preprocessed for data cleaning and visualization. After that, the OpticVib Vision Dataset (OVD), a visualized dataset of vehicle DVS, is constructed for the training and evaluation of neural networks. Finally, by improving the backbone and detection head of the network, a DVS vehicle vibration recognition and segmentation network named FiberVibNet is proposed for high-quality vehicle detection. Experimental results show that the improvements in the network structure significantly enhance the accuracy and effectiveness of FiberVibNet in detecting DVS data. Compared to the single-stage lightweight YOLO network (94.3% mAP), our model achieves superior detection accuracy (98.7% mAP) with smoother segmentation performance. Compared to Transformer-based networks, our model maintains comparable accuracy while achieving relative faster inference speed (34.0 img/s). These results demonstrate the proposed method’s capability for high-precision vehicle vibration detection.