Advanced Driver Assistance Systems (ADAS) nowadays use real-time sensors, computer vision, and embedded artificial intelligence to increase road safety. Development of affordable, multi-functional ADAS solutions for low-cost embedded platforms remains challenging due to limited computing resources and the need to process sensor data simultaneously. The purpose of this study is to present a comprehensive ADAS architecture designed for deployment on a Raspberry Pi 5 platform using the Hailo-10H AI accelerator. The proposed system integrates six safety and convenience functions: Lane Departure Warning, YOLO-based object detection, automated rearview camera activation, ultrasonic parking assistance, intelligent headlight control, and driver drowsiness monitoring based on facial landmarks and Eye Aspect Ratio estimation.
Resource allocation depends on the operational state of the vehicle, according to a multi-threaded, state-based software model. The algorithm combines object tracking with motion vector analysis to detect rear-end collisions, intersections, and side impacts. The system was evaluated using 113 simulated crash and near-miss scenarios. The proposed approach achieved a 90% detection rate for longitudinal rear events and a 60.1% overall detection rate across all scenarios tested. Simulation loop testing with BeamNG.tech further demonstrates Level 2 active intervention capabilities, including automatic emergency braking.
An Android ADAS application was created in Kotlin, targeting smartphones using TFLite inference. An automated data collection pipeline for AI training with human-in-the-loop reward scoring was created to aid future model improvement.
The obtained results demonstrate that low-cost embedded platforms can support integrated multi-sensor ADAS functions when combined with optimized state-based processing and perception, aided by artificial intelligence. Future work will focus on hardware implementation, improved training data collection, and optimization of smartphone-based inference.