This thesis develops and deploys edge-side functions for public-transport monitoring in the context of the AI-PLAYING project. The work focuses on transforming onboard sensor streams into two local indicators. The first indicator is anomaly-related and is based on reconstruction behavior in motion windows. The second indicator is a Comfort Index computed from interpretable motion terms. The prototype targets an embedded Linux platform based on an i.MX8MP System on Module and a Hailo-8L neural accelerator. The anomaly function used a Convolutional Network Autoencoder, selected to keep the reconstruction model compatible with Hailo model conversion and Hailo Runtime execution. A station-specific thresholding step converts reconstruction scores into candidate anomaly flags.