ABSTRACT

2D pose estimation is a general problem in computer vision, where the main objective is to detect a person’s body key-points and estimate a 2D skeletonized pose of a person.Skeleton estimation is outbound as an essential part of body parts detection in many fields, such as healthcare, rehabilitation, sports and fitness, animation, gaming, augmented reality, robotics. These systems are based on neural network applications and able to give reliable, objective and cost-effective benefits. Various methods are available based on this topic and used to update existing systems. In this regard, in this work, we have proposed a method for skeleton-based angle detection where we have used MobileNet model. This model is developed based on the convolution neural network (CNN). At first, 18 key-points of the human body parts were generated through the model. After that, by using the extracted key-points the skeleton of the human body parts is generated by estimating key-points according to the body part pairs. Furthermore, based on the generated skeletons, different skeleton joint angles at different key-points are estimated. To evaluate the performance of the proposed model at different environmental conditions, a customized dataset was utilized. This approach shows 95.37% accuracy for key-points detection, for joint angle estimation the accuracy is 96.11%, and shows 96.667% accuracy for body part length measurement.

Keywords: - Skeleton, Pose estimation, MobileNet, Heatmap, Convolution Neural Network.