A Probabilistic Approach based on Combination of Distance Metrics and Distribution Functions for Human Postures Classification

The article proposes a method for classifying human postures using an improved probabilistic neural network (PNN) with different distance measures and different probabilistic distribution functions (PDF). We found that the PNN with angular distance provides better accuracy, precision, and recall for the postures classification tasks than the PNN with conventional Euclidean distance. The k Nearest Neighbors (kNN) method gives slightly better prediction results than PNN, but our PNN is much faster. Such good computational performance is beneficial for posture recognition tasks that require real-time functions. An example is the needs of cobots or service robots. The article also proposes a method for selecting the distribution smoothing parameter ($\sigma$) using the sub-optimization process based on the improved Gray Wolf optimization (I-GWO) algorithm. It was found that the impact of PDF differences on the quality of the results can be reduced by choosing the best possible $\sigma$. In order to evaluate the developed method, selected human activities were recorded. The datasets were created using two different RGB-D systems located in two different laboratories.