Online Learning for Dynamic Impending Collision Prediction using FMCW Radar

Abstract

Radar collision prediction systems can play a crucial role in safety critical applications, such as autonomous vehicles and smart helmets for contact sports, by predicting an impending collision just before it will occur. Collision prediction algorithms use the velocity and range measurements provided by radar to calculate time to collision. However, radar measurements used in such systems contain significant clutter, noise, and inaccuracies which hamper reliability. Existing solutions to reduce clutter are based on static filtering methods. In this paper, we present a deep learning approach using frequency modulated continuous wave (FMCW) radar and inertial sensing that learns the environmental and user-specific conditions that lead to future collisions. We present a process of converting raw radar samples to range-Doppler matrices (RDMs) and then training a deep convolutional neural network that outputs predictions (impending collision vs. none) for any measured RDM. The system is retrained to work in dynamically changing environments and maintain prediction accuracy. We demonstrate the effectiveness of our approach of using the information from radar data to predict impending collisions in real-time via real-world experiments, and show that our method achieves an F1-score of 0.91 and outperforms a traditional approach in accuracy and adaptability.