Vehicle Path Prediction based on Radar and Vision Sensor Fusion for Safe Lane Changing

Abstract

Reported traffic accidents often occur due to rear-view blind spots. While there are many existing commercial solutions available, there is still many possible improvements. To address open issues we propose a novel approach to safe lane changing, based on radar and vision sensor fusion, which offers good accuracy with small footprint and fast performance. In the vehicle’s surrounding environment we perform deep-learning-based vehicle detection and recognition. Each vehicle is then tracked across the video sequence, with linear Kalman filter used for the spatio-temporal constraint in path prediction. Our approach achieves an accuracy of 95% in the path estimation of a vehicle approaching a blind spot.