Improved Convolutional Pose Machines for Human Pose Estimation Using Image Sensor Data

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Increasing availability of human data from image sensors necessitates efficient processing.
• Accurate human pose estimation is crucial for various applications, including robotics and surveillance.

Purpose of the Study:

• To propose a novel approach for human pose estimation by integrating Convolutional Pose Machines (CPMs) with GoogLeNet.
• To enhance the accuracy and reduce the computational cost of human pose estimation.

Main Methods:

• A hybrid model combining CPMs with GoogLeNet architecture.
• Incorporation of GoogLeNet layers into the initial stage of CPMs for improved feature extraction.
• Application of a fine-tuning strategy and inception structures to optimize the model.

Main Results:

• The improved model demonstrates superior accuracy compared to mainstream models.
• Prediction efficiency is enhanced by 1.023 times over standard CPMs.
• Training time is significantly reduced by 3.414 times due to reduced model parameters.

Conclusions:

• The proposed CPM-GoogLeNet hybrid model offers a more accurate and efficient solution for human pose estimation.
• The integration of GoogLeNet effectively addresses feature extraction limitations and reduces computational overhead.
• This research provides a promising direction for future advancements in human pose estimation techniques.