<p>In this study, in order to detect vehicles and pedestrians more effectively in thermal infrared images, we propose a multi-level feature fusion method based on YOLOv5 named YOLO-TIR. In the backbone network, we design an inverted bottleneck parallel convolution (IBPC) module, which uses wide-channel convolution in place of the deep convolutional extraction module and makes use of the Simplified SPP-Fast(SimSPPF) feature aggregation module to improve the network’s capability to localize objects. Moreover, in order to reduce the loss of information about small objects in multiple downsampling, a feature pyramid structure with jump connections called Shortcut-FPN(SFPN) is proposed, which effectively retains the shallow small object information. In addition, Shuffle-Upsample(SUP) is proposed to decrease the noise interference caused by interpolation, which splices multiple channels of the feature map to achieve the upsampling operation. Further, due to the problem of category imbalance in the public dataset, we create a new dataset called HTU_TIR to compensate for the categories with less data. The experimental results show that our proposed method has effective improvement in detection accuracy. Compared with YOLOv5m, we achieve 1.9% improvement on mAP scores, and the number of parameters and computation are decreased by 45.7% and 52.2%, respectively.</p>