INTENT: An LSTM Framework for Vehicle Intention Pre... | AI Research

INTENT: An LSTM Framework for Vehicle Intention Prediction in Intersection Scenarios with Comprehensive Ablation Analysis
Predicting the behavior of other vehicles at intersections is a critical challenge for autonomous driving. Because intersections involve complex human interactions and high-stakes maneuvers, accurately anticipating whether a car will go straight, turn left, or turn right is essential for safety and accident prevention. This paper introduces the INTENT framework, which uses Long Short-Term Memory (LSTM) neural networks to predict vehicle intentions two seconds before an action occurs, helping autonomous systems make safer, real-time decisions.

Developing the INTENT Pipeline

The researchers addressed a significant research gap: the lack of ego-vehicle datasets for intersection behaviors. To overcome this, they utilized the inD dataset, which provides bird's-eye view (BEV) trajectories of road users. Because the dataset lacked ground-truth labels for vehicle intentions, the team developed a custom pipeline. This involved a Road Labeling Algorithm to map vehicle coordinates to specific paths, a Ground Truth Labeling process to categorize maneuvers, and a Vehicle Filtration Algorithm to remove non-vehicle objects like pedestrians and cyclists. The final model processes 20-frame sequences of vehicle data, including position, heading angle, velocity, acceleration, and dimensions.

Model Architecture and Training

The core of the INTENT framework is an LSTM-based architecture designed to handle sequential trajectory data. The base model consists of three hidden layers, each containing 128 neurons and using the ReLU activation function. To ensure reliability, the team implemented several strategies:

• Data Balancing: They adjusted the input data to ensure an equal number of examples for each intention (straight, left, right), preventing the model from becoming biased toward the most common behavior.

• Optimization: The model uses the Adam optimizer with a learning rate of 0.001 and categorical cross-entropy as the loss function.

• Batch Normalization: This was applied between hidden layers to stabilize and improve the learning process.

Key Findings and Performance

The researchers conducted a comprehensive ablation study to evaluate how different configurations affected performance. By testing various feature sets, hidden layer counts, window sizes, and learning rates, they determined that their base model configuration was the most effective. The study achieved a maximum validation accuracy of 99.71%. The results demonstrate that the INTENT framework is highly effective at classifying vehicle intentions in complex intersection scenarios, providing a robust foundation for future integration into autonomous vehicle management systems.

Considerations for Future Research

While the INTENT framework shows great promise, the authors note that tuning parameters is a delicate process where changing a single factor can significantly impact results. They emphasize that while their base model performed best on the inD dataset, these findings may vary when applied to different environments or datasets. Future work aims to test the framework in diverse driving scenarios and explore its potential for broader intersection management applications.