Abstract:

This paper presents a mobile AI-based driver assistance system designed to improve pedestrian safety in the vicinity of pedestrian crossings, particularly in older vehicles that are not equipped with advanced driver assistance systems. The proposed solution uses a smartphone as an independent sensing and computing platform, combining camera-based scene analysis with GPS and inertial sensor data. The system architecture integrates pedestrian-crossing detection, pedestrian detection and tracking, short-term motion analysis, and contextual assessment of the relationship between the vehicle, the crossing, and nearby pedestrians. A dedicated relational database was developed to store synchronized visual, telemetry, trajectory, and behavioural data collected in real road conditions. These data were used to train a recurrent neural network based on gated recurrent units (GRU), which estimates a continuous driver AttentionScore subsequently mapped to six discrete AttentionLevel classes. The model processes short sequences of five consecutive observations, corresponding to approximately one second of motion history at a sampling rate of 5 Hz. In the performed experiment, the proposed model achieved a mean absolute error of 0.117, a root mean square error of 0.277, and a coefficient of determination of \(R^2=0.980\) on the test dataset. The results indicate that combining temporal pedestrian behaviour, spatial relations to pedestrian crossings, and vehicle telemetry can provide an effective basis for contextual risk assessment. The proposed architecture constitutes a low-cost edge-AI approach that may extend selected ADAS functionalities to vehicles without factory-installed vision-based safety systems.