Edge-YOLO: vehicle-mounted infrared object detection method via edge enhancement

In order to address the challenges of low image resolution, poor contrast, and complex backgrounds in infrared target detection within the field of autonomous driving, this paper proposes the Edge-YOLO detection method. This method is based on the YOLO framework and enhances detection performance through three innovative modules: ①The design of the edge convolution layer EdgeConv and the edge fusion enhancement module (EFEM), which utilize an improved Sobel operator to extract and enhance target edges while separating complex backgrounds; ②The construction of the edge perception attention module (EAM), which integrates edge information extraction, the SimAM mechanism, and feature enhancement, embedding a BottleNeck to achieve multi-scale edge feature fusion; ③The introduction of the DySample adaptive upsampling method, which optimizes parameter weights for the low-contrast characteristics of infrared images and enhances target boundary details. Additionally, the Focal EIoUloss function is employed to accelerate model convergence, with a focus on detecting small and occluded targets. The research was conducted using a self-built infrared vehicle scene dataset. Experiments demonstrated that, compared to the YOLOv8sbaseline model, the accuracy rate increased by 3.4%, the recall rate improved by 7.7%, and the mAP50 rose by4.1%, with only a slight increase in parameters and computational complexity. On the FLIR V2 dataset, the three metrics increased by 2.7%, 2.9%, and 2.4%, respectively. The proposed method effectively enhances the accuracy and efficiency of target detection in complex infrared scenes, exhibiting strong adaptability and real-time performance.