Predicting cut-in events using LSTM based techniques

Highway accidents caused by sudden cut-in and braking events pose critical challenges for driver safety, particularly when traditional Advanced Driver Assistance Systems (ADAS) only react after a cut-in is complete and time-to-collision (TTC) is critically low. This research proposes a predictive system that uses Long Short-Term Memory (LSTM) networks to forecast the probability of cut-in events and potential collisions in advance. By doing so, it enables earlier driver alerts and proactive responses, particularly in scenarios where ADAS may fail to intervene—such as when a slow-moving vehicle in an adjacent lane initiates an abrupt lane change without signaling. The model is trained using a combination of real-world driving data and synthetically generated high-risk scenarios to address data imbalance. Predictive features include relative velocity, lateral velocity, and lane overlap—key indicators of imminent risk. Results indicate an average early warning time of approximately 1.35 seconds in 40.2% of evaluated hazardous situations, significantly improving the opportunity for evasive maneuvers. At highway speeds, even a 1-second advance warning can exponentially reduce the force of impact by enabling earlier braking. The model outputs a continuous risk score ranging from 0 to 1, with continuous values in between representing the probability of a cut-in event followed by a low time-to-collision (TTC), thereby providing a quantitative measure of collision likelihood and helping prioritize driver alerts. By shifting from reactive to predictive safety strategies, this work demonstrates that LSTM-based forecasting combined with synthetic scenario modeling can enhance ADAS functionality, reduce accident rates, and contribute to the development of more intelligent, anticipatory vehicle safety systems.