Full Program »
Detecting and Predicting Smart Car Collisions In Hybrid Environments From Sensor Data
Self-driving or smart cars are becoming the next step forward for safety and autonomy of driving, aiming to provide the best travelling experience. Guaranteeing their trustworthiness requires understanding of different critical situations, such as collisions. Smart cars need to detect collisions to perform recovery actions and, even in some cases, call an ambulance. Ideally, smart cars need to predict collisions, not merely detect them after the event. This prediction is difficult, especially when human drivers are part of the environment. This work proposes a system for collision prediction and detection based on machine learning: it focuses only on the information that the car receives through its sensors. We evaluate our method using a multi-car hybrid system designed to generate realistic and complex scenarios for the smart cars, thereby providing sensory information that is close to a realistic road. Our methodology detects collisions with 99% accuracy and predicts 5 different levels of collision risk with up to 43% accuracy.