Research on infrared small target pedestrian and vehicle detection algorithm based on multi-scale feature fusion

Infrared imaging technology relies on detecting the electromagnetic waves emitted by an object's spontaneous thermal radiation for imaging. It can overcome the adverse effects of complex lighting conditions on the detection of pedestrians and vehicles on the road. To address the issues of low accuracy and missed detection in visual detection under complex traffic conditions, such as during rain, snow, or at night, a pedestrian and vehicle detection model using infrared imaging has been proposed. This model improves the neck network and incorporates an attention mechanism. First, by adding a multi-scale feature fusion small-object detection layer to the model's neck, enhancing the capture of detailed information about small infrared objects and reducing missed detections. Second, a novel dual-layer routing attention mechanism is designed, allowing the model to focus on the most relevant feature areas and improving the detection accuracy of small infrared objects. Next, the CARAFE upsampling method is used for adaptive upsampling and context information fusion, which enhances the model's ability to reorganize features and capture details. Finally, a lightweight CSPPC module is constructed using partial convolutions to replace the C2f module in the neck network, which improves the model's frame rate. Experimental results show that, compared to the baseline model, BCC-YOLOv8n improves precision, recall, mAP@0.5, and mAP@0.5:0.95 by 1.4%, 4.8%, 5.3%, and 4.5%, respectively, while reducing the number of parameters by approximately 7%. Additionally, a frame rate of 70.8 FPS was achieved, satisfying the requirements for real-time detection.