O-LiPE: An Optimized Lightweight Human Pose Estimation System with Structured Pruning and Quantization for Real-Time Inference

Abstract

The use of human poses in estimating human poses has gained tremendous importance in a number of computer vision systems, such as activity recognition and human-computer interaction, as well as intelligent surveillance systems. However, many state-of-the-art pose estimation models rely on complex deep neural networks with large computational requirements, making them difficult to deploy in real-time or resource-constrained environments. In order to overcome this constraint, this paper presents a light human pose estimation model that emphasizes enhancing computational efficiency with an acceptable prediction ability.The suggested method is an implementation of a backbone based on MobileNetV2 and depthwise separable deconvolution layers to minimize the model parameters. Besides, an attention refinement mechanism is included to improve the description of spatial features. In order to optimize the model further to be used in a real-time system, structured pruning and dynamic quantization techniques are used. The structured pruning operation eliminates all of the redundant channels in the network, thus significantly reducing the number of parameters, and quantization reduces the numeric accuracy of model weights to reduce memory usage.The model is trained using the COCO keypoint dataset with heatmap-based supervision and evaluated through both qualitative visualization and computational performance analysis. The experimental findings indicate that the optimized model can realize a significant reduction in the model size and attain real-time inference. The optimised and quantised network has around 0.36 million parameters and has an inference rate of around 55 frames a second. These findings suggest that the suggested framework can be used as a practical solution when it comes to lightweight human pose estimation in real-time scenarios.