Abstract: In-vehicle infrared images can help drivers identify pedestrians and other vehicles on the road at night and during bad weather, thereby reducing traffic accidents. To address the low detection accuracy of vehicle infrared images using the YOLOv5s algorithm, an improved YOLOv5s algorithm for vehicle infrared image target detection is proposed. First, a receptive field enhancement structure, namely an RFENeck module, is designed, which replaces the BottleNeck module in C3, to enhance the receptive field area of the feature fusion network and thus improve the detection accuracy. Second, a dynamic object detection head, combined with an attention mechanism, is used to improve the expression ability of the detection head. Finally, to eliminate the increase in model size caused by the improvement, an efficient backward residual mobile module, combined with the cascade designs of convolutional neural networks and Transformer, is used to form the backbone network. This module can reduce the number of network parameters and calculation steps without reducing the accuracy. The experimental results show that compared to YOLOv5s, the average detection accuracy of the improved algorithm increases from 82.9% to 85.0%; in addition, the computation amount is reduced by 5.7%, and the model weight is reduced by 0.4 M. These results indicate that the proposed algorithm fulfils the requirements of model size and accuracy.