Highway on-ramp merging remains one of the most challenging maneuvers for both human drivers and autonomous systems. This research, presented at AAAI-20, proposes a collaborative approach that shifts the computational burden of intention prediction from individual vehicles to roadside infrastructure. By utilizing Road Side Units (RSUs) to aggregate traffic data, the system achieves high-accuracy predictions of driver intent several seconds before a merge occurs.
The Case for Off-Board Intelligence
Traditional autonomous vehicles (AVs) rely heavily on on-board sensors like LiDAR, radar, and cameras. While powerful, these sensors are limited by line-of-sight obstructions and the localized perspective of a single car. In contrast, an "off-board" solution uses RSUs to maintain a holistic view of the traffic environment. This centralized perspective allows for better coordination between merging vehicles and those already in the main highway lanes, potentially leading to safer and more efficient traffic flow.
A Hierarchical Model for Driver Intent
The researchers developed a three-layer hierarchical model to translate raw sensor data into actionable predictions:
- Vehicle State (Vector X): Captures raw dynamic data such as position, speed, and acceleration.
- Situational Context (Vector C): This relational layer calculates the physics between key participants: the merging vehicle, the vehicle directly ahead in the merge lane, and vehicles in the main highway lane.
- Intention (Vector I): The final output—a binary prediction of whether the vehicle will merge.
To process this data, the study compared discriminative classifiers (Logistic Regression, k-Nearest Neighbors, and Neural Networks) against generative models (Naïve-Bayes and its variants).
Performance and Predictive Windows
Using the Next Generation Simulation (NGSIM) dataset from Interstate 80 in California, the researchers found that discriminative models significantly outperformed generative ones. Logistic Regression (LRM) achieved a remarkable 99% accuracy, precision, and recall.
Crucially, the model provides a significant lead time. It identifies a driver's intention approximately four seconds before the actual merge point. Given that a typical connected vehicle system requires only about 0.5 seconds for mechanical response and network latency, this four-second window provides a substantial safety buffer for autonomous decision-making.
Future Directions
While the results are promising, the study highlights the need for extensive infrastructure—specifically a high density of RSUs—to make such a system viable. Future work will likely explore hybrid systems that combine on-board sensing with off-board coordination, as well as testing the model's robustness across diverse weather conditions and driving environments.
Summary updated