Multifeature-fusion-based Vehicle Reidentification for Tunnel Scenes

To solve the problem of low reidentification accuracy caused by the low resolution and uneven illumination of traffic-monitoring video in tunnels, a multifeature fusion method for vehicle reidentification is proposed to improve the accuracy by making full use of various types of feature information about vehicles. First, a convolutional block attention module was embedded into the convolutional layer of the backbone network of the YOLOv5 model, while complete intersection-over-union loss and distance intersection-over-union nonmaximum suppression schemes were used to improve the vehicle model detection accuracy. Second, an improved DeepSORT apparent feature extraction network and ResNet network were used to obtain deep convolutional neural network (DCNN) features and vehicle ID features, respectively. Third, model, DCNN, and vehicle ID features were fused using the sum representation layer to form identifiable identity features to improve reidentification accuracy. Finally, the indicator function was designed, and learning was performed based on softmax cross-entropy and triplet losses. The learning results were reordered to improve the accuracy of the model further. The algorithm was trained and validated on the public datasets VeRi776 and VehicleID and the self-built Tunnel_Veh4C dataset. Compared with previous methods, the recognition accuracies of Rank 1, Rank 5, and Rank 10, as well as the mean average precision (mAP), were improved to different degrees on the datasets. The mAP increased by 28.78%, 20.68%, 18.05%, and 11.99%, while the Rank 1, Rank 5, and Rank 10 recognition accuracies reached 95.12%, 98.48%, and 99.3%, respectively, on Tunnel_Veh4C. In addition, the proposed method was applied to engineering projects, and good results were achieved. Therefore, this method can improve vehicle reidentification accuracy by fully using the various types of vehicle feature information. It has good reidentification performance for vehicles under different viewpoints and lighting, providing strong robustness in practical applications. © 2023 Xi'an Highway University. All rights reserved.