Harsh Event Detection in Sensor Fusion Systems with Improved Event Filtering and Noise Removal Techniques
Abstract
: Accident detection and prevention has been one of the key areas of research in the domain of Advanced Driving Assistance Systems and self-driving cars, as it is directly associated with loss of life and property. Most crashes occur due to harsh braking at high speeds, unsafe lane changes, turns, intense acceleration, and traffic rule violations. ADAS (Advanced Driver Assistance Systems) are installed on vehicles to detect such events. Still, they rely heavily on a suite of expensive sensors like radars, LIDARs (Light Detection And Ranging), and mounted cameras for collision detection and warnings. In this research, a cheaper and effective framework for harsh event detection is proposed. It relies only on telematics data generated from a set of inexpensive sensors like the IMU (Inertial Measurement Unit), GPS (Global Positioning System), and vehicle CAN (Controller Area Network). The proposed system involved generating harsh event candidates for Machine Learning (ML) model training, optimization, and testing on a set of benchmark datasets to evaluate its performance in different scenarios. The results revealed that XGB (Extreme Gradient Boosting) outperformed other applied ML techniques, such as NB (Naïve Bayes), SVM (Support Vector Machine) and DT (Decision Tree), with 98% precision and 99% recall for the test dataset, 90% precision and 88% recall for the private industry dataset, and 82% precision and 90% recall for the public literature dataset.