Improving infrared road object detection with YOLOv8

To address the challenge of insufficient detection rates and precision for miniature objects in infrared road scenes, we propose a refined object detection algorithm called SSTD-YOLO built upon YOLOv8 tailored for infrared road environments. Our approach begins by leveraging SPD-Conv to mitigate the loss of detailed information stemming from the prevalence of small objects and low image resolution. Furthermore, we integrate a triplet attention mechanism within the C2F module to enhance the identification of complex samples. Additionally, our model incorporates Dysample in the feature fusion module, surpassing alternative upsampling techniques in inference latency, memory consumption, FLOPs, and parameter count. To augment the detection accuracy of small targets in infrared images and alleviate instances of undetected small targets, dedicated layers for small target detection are integrated into our model. Experimental results demonstrate that our proposed model outperforms the YOLOv8n baseline, increasing mAP@0.5 by up to 14.16%.