An original spillage risk vehicle detection method called spillage risk vehicles detection network (SRVDNet) was proposed faced with the challenge of detecting spillage risk vehicles, specifically designed for high-speed freight, to resolve the issues which insufficient extraction of spillage risk vehicle features and inadequate fusion of deep features in existing methods. A backbone network incorporating large separable kernel was introduced to enhance the network’s learning capabilities for spillage risk features in freight vehicles. A neck network integrating feature gather-and-distribute mechanisms was introduced to provide the detection head with richer information such as freight carriage types, tarpaulin edge textures and cargo shape contours. The proposed method was validated using a real high-speed highway freight vehicle dataset. The experimental results demonstrate that SRVDNet exhibits superior performance, achieving a detection accuracy of 81.5%. SRVDNet shows improvements of 3.70%, 3.09%, 2.86%, 1.37%, 1.41%, 2.00%, respectively, in terms of mAP@0.5 metrics compared with existing state-of-the-art (SOTA) object detection algorithms such as YOLOv5, YOLOv6, YOLOv8, RT-DETR, PP-YOLOE, YOLOv9. The model parameters were relatively smaller, while detection speed remains high. This method effectively enhances the ability to identify spillage risks in scenarios with irregular cargo loading, low cargo volume, and full cargo loads, thereby contributing to source control of spills and strengthening the capability for early warning of safety risks on highways.