Pedestrian Behavior Analysis and Group Detection: An Integrative Clustering and Dynamic Graph Framework

In recent years, understanding pedestrian behavior has become crucial due to advancements. Although much research has been done on clustering trajectories, no one has looked at using pedestrian datasets to identify patterns from movement and categorize behaviors in an unsupervised manner. To address this gap, the first part of this research develops an unsupervised framework using an LSTM-based autoencoder to identify the distinct walking behaviors of pedestrians. Unlike previous studies that rely on labeled datasets, our unsupervised approach eliminates the need for manually labeling pedestrian data. This model can categorize behaviors in complex environments like walkways, public squares, or shopping centers by learning the underlying patterns in pedestrian movements by analyzing the sequences of their walking paths. The second part enhances the study by exploring bottlenecks in pedestrian movement through graph-based analysis. This dynamic graph-based methodology allows us to visualize and measure areas where pedestrian interactions intensify or slow down. This insight is crucial for modelling pedestrian flow in shared spaces, aiding in designing more walkable environments and efficient emergency evacuation strategies. The model was validated using UCY dataset, a well-known benchmark for pedestrian trajectory data and its performance was evaluated using key performance metrics such as accuracy, precision, and recall. The final phase involved developing a group detection mechanism integrating behavior classification and graph-based analysis. This method is particularly effective for surveillance by preserving the privacy of individuals - no personal or image data from a surveillance camera is needed. Overall, the framework and developed mechanism are adaptable to numerous scenarios, contributing to improved pedestrian safety, walkability, emergency evacuation planning, and pedestrian behavior analysis in shared spaces.