Lane-changing is a routine but difficult driving task with important implications on traffic flow characteristics. Despite significant progresses in lane-changing decision modeling, lane-changing models are often improperly calibrated due to two issues related to trajectory data processing. First, the time of insertion (i.e., the time instant where a vehicle crossed the lane marking) is incorrectly considered as the lane-changing decision point, since the lane-changing decision is typically made earlier. Secondly, there is an imbalance between the number of non-lane-changing and lane-changing events, where non-lane-changing events typically dominate trajectory data. These issues can overestimate model performance and biased parameters. In this paper, we propose a method that combines (i) the wavelet transform method to pinpoint the correct lane-changing decision point, and (ii) a case–control design to systematically neutralize the dominance of non-lane-changing events in the data. The proposed method is applied to two NGSIM datasets to assess the performance of four representative lane-changing models. Results uncover that (i) lane-changing models are sensitive to various degrees of data imbalance, (ii) regardless of a driver’s decision time window (e.g., 1 s, 2 s, or 3 s), an analysis time window of 6 s will work reasonably well for evaluating the performance of a lane-changing model, while the optimal control-to-case ratio is 1:1; and (iii) when possible, a temporal discretization interval (i.e., an approximation of a driver’s typical decision time window) of 2 s should be preferred, while 3 s should be avoided. The proposed method also enabled us to outline a performance range for the selected lane-changing models.
Read full abstract