Investigating the dynamic memory effect of human drivers via ON-LSTM

It is a widely accepted view that considering the memory effects of historical information (driving operations) is beneficial for vehicle trajectory prediction models to improve prediction accuracy. However, many commonly used models (e.g., long short-term memory, LSTM) can only implicitly simulate memory effects, but lack effective mechanisms to capture memory effects from sequence data and estimate their effective time range (ETR). This shortage makes it hard to dynamically configure the most suitable length of used historical information according to the current driving behavior, which harms the good understanding of vehicle motion. To address this problem, we propose a modified trajectory prediction model based on ordered neuron LSTM (ON-LSTM). We demonstrate the feasibility of ETR estimation based on ON-LSTM and propose an ETR estimation method. We estimate the ETR of driving fluctuations and lane change operations on the NGSIM I-80 dataset. The experiment results prove that the proposed method can well capture the memory effects during trajectory prediction. Moreover, the estimated ETR values are in agreement with our intuitions.