Introduction
Modern Advanced Driver Assistance Systems (ADAS) have significantly improved road safety through reactive measures like emergency braking and lane-keeping assistance. However, these systems typically intervene only when a collision is imminent. To further reduce highway fatalities, there is a need for proactive systems capable of predicting dangerous maneuvers before they occur. This paper proposes a predictive framework that models driving behavior as a stochastic process using Hidden Markov Models (HMMs) to anticipate crashes and trigger preventive actions.
Problem Statement
Highway environments are inherently dangerous due to the unpredictable nature of human drivers and complex interactions between vehicles. The primary challenge in proactive safety is intent prediction: accurately estimating a vehicle's future trajectory and speed based on its current state. Existing systems often lack the temporal foresight to distinguish between different types of collision threats, such as rear-end versus side-swipe accidents, which require different evasive maneuvers.
Methodology
The proposed system utilizes a two-layer HMM architecture to model vehicle behavior. By analyzing historical trajectory data from the USDOT Intelligent Transportation Systems (ITS) database, the model estimates the probabilities of lane changes and speed variations.
Two-Layer HMM Architecture
The model is divided into two distinct layers to maintain computational efficiency while providing granular predictions:
- Layer 1 (Lane Change): Represents the vehicle's position across six highway lanes.
- Layer 2 (Speed Change): Represents the vehicle's speed across six discrete ranges, from 0 to 60 m/s.
The transition between these states is modeled as a Markov process, where the probability of moving to a new state depends only on the current state and the transition probabilities derived from historical data.
Real-Time Integration and Decision Logic
The system integrates real-time data from Lidar and Vehicle-to-Vehicle (V2V) communication. Lidar provides precise measurements of relative distance ($d$) and relative velocity ($V$), allowing the calculation of a "probable time of crash" ($t = d/V$).
A key innovation is the use of the Mean First Passage Time (M), a metric that calculates the expected time for a vehicle to transition from its current lane to a target lane. The decision logic follows these steps:
- If the predicted crash probability exceeds a threshold (0.3), the system compares $t$ with $M$.
- Rear-end Threat ($M > t$): If the expected lane-change time is greater than the crash time, the vehicle is likely to remain in its lane. The system activates Adaptive Cruise Control (ACC) to mitigate a rear-end collision.
- Side-swipe Threat ($M < t$): If the vehicle is likely to change lanes before the crash time, the system triggers steering assistance to prevent a sideways collision.
Experimental Results
The framework was validated using simulated scenarios in Matlab/Simulink based on instrumented vehicle trajectory data.
- Scenario 1 (Rear-end): The model predicted a 76.78% crash probability. Since $M > t$, the system correctly identified a longitudinal threat and engaged ACC for the following car.
- Scenario 2 (Side-swipe): With a 75.36% crash probability and $M < t$, the system identified a lateral threat and successfully triggered steering assist to keep the vehicle in its lane.
- Scenario 3 (Safe Operation): In standard driving conditions with adequate spacing, crash probabilities remained below 5%, resulting in no unnecessary interventions.
Conclusion and Future Work
This research demonstrates that a layered HMM approach can effectively predict highway crashes by treating driving as a probabilistic sequence of states. By incorporating the Mean First Passage Time, the system can intelligently choose between braking and steering interventions. Future work will focus on increasing model granularity and accounting for human factors such as braking intensity and environmental conditions to further refine the accuracy of crash avoidance.
Summary updated