A geometry-driven car-following distance estimation algorithm robust to road slopes

摘要

Locating the surrounding vehicles is an important environment perception task for autonomous vehicles and advanced driver assistance systems. This task is usually explored based on the sensors' pre-calibration (e.g. height or pitch angle), but can be challenging when the calibration fails (e.g. on the sloping and uneven roads). In this work, we propose a calibrated feature-point based (CFPB) method to estimate the car-following distance adaptive to rough roads, using a single camera. Instead of using the pre-calibrated parameters, CFPB method is based on the surrounding vehicles' feature points. It benefits from the fixed coordinate relations among these points, which enables the algorithm to be adaptive to rough roads. These fixed coordinate relations can be calibrated during driving. Namely, after a few seconds of observation to a surrounding vehicle, the CFPB method can start working for more accurate estimation. Furthermore, the proposed algorithm takes the perspective-n-point method as the framework. YOLO V3 and scale-invariant feature transform are applied as the vehicle detector and feature point extractor. The feature point calibration is dynamically updated and the results are smoothened by a Kalman filter. The camera is chosen because of the good performance on objects detection and feature extraction. The proposed algorithm is evaluated on a real-world road with dynamic traffic flow. Mobileye, a widely used car-following distance estimator on AV, is installed during the tests as the benchmark. The results indicate that the proposed method achieves decimeter-level accuracy and outperforms the Mobileye system in cases where the road slope changes significantly.