[Paper Review] Identifying Modes of Intent from Driver Behaviors in Dynamic Environments

Katherine Driggs-Campbell, Růžena Bajcsy|arXiv (Cornell University)|May 21, 2015

Autonomous Vehicle Technology and Safety11 references4 citations

TL;DR

This paper proposes a hybrid system model that identifies driver intent in dynamic driving environments using real-time, human-labeled lane change data from a driving simulator. By classifying intent based on environmental sensor features rather than control inputs, the model achieves up to 92.1% accuracy in identifying intent modes, with SVM showing the most robust and smooth transitions.

ABSTRACT

In light of growing attention of intelligent vehicle systems, we propose developing a driver model that uses a hybrid system formulation to capture the intent of the driver. This model hopes to capture human driving behavior in a way that can be utilized by semi- and fully autonomous systems in heterogeneous environments. We consider a discrete set of high level goals or intent modes, that is designed to encompass the decision making process of the human. A driver model is derived using a dataset of lane changes collected in a realistic driving simulator, in which the driver actively labels data to give us insight into her intent. By building the labeled dataset, we are able to utilize classification tools to build the driver model using features of based on her perception of the environment, and achieve high accuracy in identifying driver intent. Multiple algorithms are presented and compared on the dataset, and a comparison of the varying behaviors between drivers is drawn. Using this modeling methodology, we present a model that can be used to assess driver behaviors and to develop human-inspired safety metrics that can be utilized in intelligent vehicular systems.

Motivation & Objective

• To develop a portable, human-centered driver model that captures high-level intent in dynamic, heterogeneous traffic environments.
• To model driver decision-making using a hybrid system formulation where mode transitions are driven by environmental dynamics rather than time or heuristics.
• To create a robust classification framework that uses real-time human-labeled intent data to train models based on environmental perception features.
• To enable application in semi- and fully autonomous systems by ensuring the model is independent of driver state or control inputs.
• To establish a foundation for human-inspired safety metrics and high-level decision-making in mixed-traffic autonomous systems.

Proposed method

• Collected a real-time, human-labeled dataset of lane changes in a high-fidelity driving simulator, with drivers explicitly labeling their intent during maneuvers.
• Formulated driver intent as a hybrid system with discrete modes (e.g., lane keeping, lane changing), where mode transitions are triggered by dynamic environmental states.
• Extracted perception-based features from surrounding vehicles (e.g., relative position, speed, distance) as inputs to the classification model.
• Trained and compared multiple supervised classifiers—SVM, Random Forest (RF), and Logistic Regression (LR)—on the labeled dataset to detect intent modes.
• Evaluated models using overall accuracy and scenario-specific accuracy (based on number of nearby vehicles), with normalization to reduce class imbalance bias.
• Used confusion matrices and normalized probability outputs to analyze classification reliability and transition smoothness across modes.

Experimental results

Research questions

• RQ1 Can driver intent in dynamic environments be accurately identified using perception-based features rather than control inputs or driver state?
• RQ2 How do different classification algorithms (SVM, RF, LR) perform in detecting mode transitions during lane changes based on environmental dynamics?
• RQ3 To what extent does the model’s accuracy vary across drivers, and what factors contribute to inter-driver variability in intent labeling?
• RQ4 Can a hybrid system model that uses environment-driven mode transitions outperform traditional time-windowed or heuristic-based intent detection methods?
• RQ5 Is the model robust and generalizable enough to serve as a high-level controller for autonomous vehicles in mixed-traffic environments?

Key findings

• The SVM classifier achieved the highest overall accuracy of 89.5% and demonstrated the smoothest mode transitions, with a conservative but reliable prediction of lane changes before execution.
• Random Forest (RF) achieved 88.9% overall accuracy and showed strong mode discrimination, though with less smooth transitions due to fine-grained boundary detection.
• Logistic Regression (LR) produced lower overall accuracy (87.2%) but effectively captured probabilistic mode transitions, revealing uncertainty during maneuver boundaries.
• Accuracy varied significantly across individual drivers, with the highest-scoring scenario (Scen. 3, 3 nearby vehicles) reaching 92.1% for SVM, indicating environmental complexity affects performance.
• Combining data from all drivers to build a global model resulted in extremely poor performance, indicating that individual driver modeling is essential for reliable intent detection.
• The model achieved strong results without using windowing techniques, demonstrating that minimal feature preprocessing maintains flexibility and generalization across diverse driving conditions.