Target vehicle lane-change intention detection: An approach based on online transfer learning

The lack of visual information from the driver of the target vehicle makes it difficult to detect the intention of the target vehicle early before the start of the lane-change maneuver. Moreover, in current studies, the prerequisite for the successful application of intention detection models is to obtain a sufficient number of lane-change maneuver samples covering various scenarios and characteristics. Thus, a universal detector may not be optimal for lane-change intention detection in different scenarios in the real world. In this paper, we propose an intent detection method based on online transfer learning (OTL). First, a passive-aggressive (PA) algorithm is adopted to construct a lane-change source classifier with a large number of lane-change maneuvers as training samples. These samples contain the motion parameters of the vehicle before changing lanes, and the relative motion relationship between the target vehicle and the surrounding vehicles. Then, an OTL strategy that automatically and dynamically updates the weights is designed to detect the intention of the target vehicle to change lanes. The construction of the source classifier is supported by Next Generation Simulation (NGSIM) natural driving data, and the verification of the intention predictions exploits data collected from real natural driving experiments. The performance analysis results demonstrate that the proposed method can successfully detect lane-change intention 3 s before the start of the lane-change maneuver, with an accuracy of 93.0%.