Assessing risk level of lane changing behaviors using time series analysis

Dangerous lane-changing maneuvers substantially increase the risk of collisions of individual vehicles on urban roads. To improve the driving safety of vehicles, it is very important to differentiate risky and safe driving behaviors. This research harnesses an advanced Long Short-Term Memory (LSTM) model to classify and discern potential high-risk and aggressive lane changes. The study first conducts an exhaustive risk evaluation, attributing a risk score to each lane change case, based on real-world data. A k-means clustering of these risk scores is undertaken to meticulously categorize each lane change instance into two distinct groups: “Not Dangerous” and “Dangerous.” An LSTM model is improved with time-series data to predict lane-changing risk levels corresponding to the labeled risk categories. To bolster the model’s performance on imbalanced data, the LSTM model is augmented with the Synthetic Minority Over-sampling Technique (SMOTE) to regenerate data concerning aggressive lane changes. This technique is leveraged to reproduce data related to aggressive lane changes, thus successfully rectifying data imbalance issues and enhancing prediction accuracy. It is observed that the proposed method demonstrated superior performance when compared to other prevalent classification methods, including Support Vector Classification (SVC) and decision tree algorithms. This research encourages a more proactive risk identification process, potentially paving the way for the development of safer and more intelligent driving systems.