Deep and Classical Machine Learning for Skeleton-Based Gait Stance Recognition

There has been a noticeable upsurge in the amount of research conducted in the last several years on the analysis of human behavior, especially among computer vision experts. Technological developments and the growing uses of human activity recognition (HAR) in various industries, including healthcare, sports analytics, and human-computer interface, are responsible for the growing interest. Deep learning techniques are the most promising of the different approaches to HAR because of their capacity to extract intricate patterns from enormous datasets. In this paper, we present a methodology for characterizing human activities utilizing 3D skeleton information. A wide range of human movements can be accurately recognized and categorized using this method, which makes use of deep learning's capabilities. The center parts of our approach are illustrated as Joint Position Scaling and Temporal Alignment. After normalization, the 3D skeleton data is represented in two different forms like Spatial Representation and Temporal Representation. Support Vector Machine (SVM) has been employed on the features extracted from the deep neural network to improve our model's classification accuracy. The output from the second last layer of the neural network is used to train the SVM, allowing it to draw a decision boundary that effectively divides movements of gait. Bayesian Optimization technique has been used to adjust each hyperparameter to make sure our model was resilient. This approach is very useful as it finds the best configurations by quickly exploring the hyperparameter space. The PKU-MMD dataset, which offered a thorough baseline for our model's training, was used for the optimization.