Efficient high-resolution human pose estimation with integrated attention mechanism

Deep learning has brought about a revolutionary transformation in the field of human pose estimation, offering significant advantages. Traditional approaches often seek performance improvement by expanding and deepening networks, resulting in increased parameters and complexity. In response to this challenge, we introduce DNNet, a novel framework rooted in HRNet. Its unique basic blocks incorporate feature mapping, channel weighting, and the fusion of different receptive fields. In comparison to HRNet, DNNet stands out with fewer parameters, lower computational complexity, and more resilient lightweight features. By leveraging features at various resolutions, DNNet's performance is further enhanced. Extensive validation on both the COCO dataset and a dataset focused on dangerous driving behavior reveals that DNNet's accuracy is comparable to that of HRNet. Notably, DNNet achieves a remarkable 69% reduction in parameters and a 59% decrease in computational complexity under similar accuracy conditions. In practical real-world applications, when applied to a dataset addressing dangerous behavior, DNNet outperforms both ShuffleNet and MobileNet in accuracy, highlighting its adaptability and efficacy in diverse scenarios. DNNet's unique features position it as a promising solution in the field of human pose estimation, providing a balanced trade-off between accuracy, efficiency, practical applicability, and the organic fusion of different receptive fields.