Detailed Analysis of Pedestrian Activity Recognition in Pedestrian Zones Using 3D Skeleton Joints Using LSTM

As autonomous driving technology is developing rapidly, demands for pedestrian safety, intelligence, and stability are increasing. In this situation, there is a need to discern pedestrian location and action, such as crossing or standing, in dynamic and uncertain contexts. The success of autonomous driving for pedestrian zones depends heavily on its capacity to distinguish between safe and unsafe pedestrians. The vehicles must first recognize the pedestrian, then their body movements, and understand the meaning of their actions before responding appropriately. This article presents a detailed explanation of the architecture for 3D pedestrian activity recognition using recurrent neural networks (RNN). A custom dataset was created for behaviors such as parallel and perpendicular crossing while texting or calling encountered around autonomous vehicles. A model similar to Long-Short Term Memory (LSMT) has been used for different experiments. As a result, it is revealed that the models trained independently on upper and lower body data produced better classification than the one trained on whole body skeleton data. An accuracy of 97% has been achieved for lower body and 88–90% on upper body test data, respectively.