Improving human pose recognition with YOLOv8-SLG using thermal images

Abstract

Thermal imaging cameras operate effectively at night or in low-light conditions, enhancing the accuracy of human pose recognition in such environments. To address the issue of insufficient accuracy in traditional human pose recognition methods under low-light conditions, this study proposes an improved algorithm based on the YOLOv8 model, referred to as YOLOv8-SLG. Considering the peculiarity that thermal images reflect temperature information rather than optical features, traditional convolutional networks may produce unnecessary redundancy in feature extraction. To address this issue, first, we improve the detection accuracy and preserve the network structure by replacing the original convolution with SCConv building blocks in the backbone network of YOLOv8n in order to reduce spatial and channel redundancy between features in the convolutional neural network. Second, we enhance local feature detection by integrating the LSKA attention mechanism into the neck network, reducing computational complexity and memory requirements while maintaining accuracy. Finally, we enhance the multi-scale processing capability and reduce the number of parameters per detection head through shared GroupNorm convolution to improve target localisation and classification performance. Experimental results show that these enhancements significantly improve the model’s performance for human pose recognition tasks in complex contexts. Compared to the original YOLOv8n model, the proposed algorithm improves the precision, recall, mAP50, and mAP50-95 metrics by 1.33%, 1.79%, 1.86%, and 2.01% to 97.7%, 94.6%, 96.7%, and 75.1%, respectively. In addition, YOLOv8-SLG reduced model parameter calculations by 8.14%. It can detect human poses in thermal images in real time accurately, and comparison with other mainstream human pose detection algorithms confirms the effectiveness and superiority of the method.