Lane Change Intention Inference of Surrounding Vehicle: Comparative Study on Relevance Vector Machine (RVM) and Support Vector Machine (SVM)

Abstract

This paper presents a methodology to infer the intention to lane change of Surrounding Vehicle (SV) by designing a classifier using a Relevance Vector Machine (RVM). Estimating the intentions precisely of SV is one of the key technologies in autonomous driving. In particular, the lane change of SV is a situation that can be frequently observed while driving, and the behavior of the host vehicle may be affected by the SVs. Therefore, we propose a probabilistic learning-based intention classifier and conduct a performance validation. The training data was reshaped by processing the sensor data acquired from the actual driving and extracting the characteristic points of the maneuver. The verification data was collected on a road not included in the training data. We conducted a comparative experiment with the RAdio Detecting And Ranging (RADAR) signal for the front target and a deterministic classifier called a Support Vector Machine (SVM). Statistically, the proposed RVM-based method succeeded in predicting the lane change of the surrounding vehicle faster than both the RADAR sensor and the SVM. In addition, we observed that correct intention classification was performed through RVM even if SVM misclassified the class set of a lane change by some irregular lateral motion of SV.