Factory Small Target Pose Estimation Using Transformer

Addressing the limitations of the YOLOv8-Pose pose estimation algorithm in terms of accuracy and effectiveness for small target detection tasks in industrial scenarios, this paper proposes a small target human pose estimation algorithm that incorporates a Transformer. The aim is to enhance the accuracy of the original algorithm and improve its recognition performance in specific scenarios. Firstly, an improved ODCoT module is proposed by introducing an enhanced full-dimensional dynamic convolution into the CoT module. This improved module captures the correlation of internal features and leverages the global modeling capabilities of the Transformer to extract small target features more accurately. Secondly, the PAN et structure is refined to fully integrate shallow and deep features, thereby enhancing the network's sensitivity to small targets. Finally, the loss function strategy is adjusted to accelerate model convergence and improve precision. Experimental results indicate that the improved algorithm achieves an increase in mean accuracy compared to algorithms such as YOLO-Pose and YOLOv8-Pose on a self-made dataset of small human targets in a factory setting.