MNCL: Empirical Development of Human Action Recognition Based on Postures Using Modified Neural Classification Logic

In this research, we introduce an innovative method for the recognition of human actions based on skeletal joint data, which offers greater privacy compared to color photos. The methodology includes defining an activity using a sequence of instructive fundamental postures, and combining machine learning methods for activity recognition. Based on normalized skeletal data points, posture features such as absolute joint positions, relative joint positions, and joint angle quaternions are proposed to derive. Feature dimensionality and system complexity are reduced through posture selection, followed by temporal modeling to minimize the effects of rate variation, temporal misalignment and noise. A Modified Neural Classification Logic (MNCL) model is proposed, based on certain principles, to identify actions. Position- and angle-based kinematics are suggested, with angle based features that are naturally normalized and are independent of an individual's physique. To reduce posture similarity and make the method more general, a posture selection method based on the K-means clustering algorithm is implemented. The proposed MNCL model links weights of the MNCL model between the previous layer and the current layer using a linear function instead of the non-linearity provided in the original MNCL model. An innovative inference engine for logical connections based decision making is used to improve the output layer recognition rate. In this work, we propose a system for human action recognition based on skeletal joint properties, intended to provide an efficient and privacy preserving solution.