Andrii Bidochko

• Updated: July 12, 2026
• 7 min read

INTENT: An LSTM Framework for Vehicle Intention Prediction in Intersection Scenarios

Direct Answer

The paper introduces INTENT, an LSTM‑based framework that predicts a vehicle’s maneuver (straight, left turn, or right turn) two seconds before it occurs at intersections. By achieving 99.71 % accuracy on the InD dataset, INTENT demonstrates that deep sequential models can reliably infer driver intent in complex, high‑risk traffic scenarios, a capability that directly supports safer autonomous‑driving systems.

Background: Why This Problem Is Hard

Intersection zones are among the most unpredictable environments for autonomous vehicles. Drivers must interpret a mixture of visual cues—traffic lights, road markings, surrounding vehicle behavior, and even subtle driver gestures—to decide whether to proceed straight, turn left, or turn right. Traditional perception pipelines excel at detecting objects but often treat intention as a downstream classification problem, leading to two critical bottlenecks:

• Temporal ambiguity: The observable cues that signal a driver’s future action can appear only moments before the maneuver, leaving little reaction time for an autonomous system.
• Contextual complexity: Intersections involve multiple agents, varying traffic rules across jurisdictions, and occasional emergency maneuvers (e.g., sudden stops), which strain rule‑based or shallow‑learning models.

Existing approaches—such as handcrafted rule sets, Bayesian inference, or shallow neural networks—struggle to capture the long‑range dependencies and nuanced patterns embedded in sequential sensor data. Consequently, they either produce high false‑positive rates or miss critical early warnings, compromising both safety and traffic flow efficiency.

What the Researchers Propose

The authors present INTENT, a framework that leverages a Long Short‑Term Memory (LSTM) network to model the temporal evolution of vehicle dynamics and surrounding context. The core idea is to treat intention prediction as a sequence‑to‑label problem: given a sliding window of sensor observations, the LSTM outputs a probability distribution over three possible maneuvers.

Key components of the framework include:

• Feature encoder: Normalizes and concatenates multi‑modal inputs (e.g., position, velocity, yaw rate, traffic‑light state) into a fixed‑size vector per timestep.
• LSTM core: Captures long‑range temporal dependencies across the observation window, allowing the model to “remember” early cues that influence later decisions.
• Intention classifier: A dense layer with softmax activation that translates the final hidden state into the three‑class intention output.

By focusing on a 2‑second prediction horizon, INTENT balances the need for early warning with the practical limits of sensor latency and computational budget, making it suitable for real‑time deployment on automotive edge hardware.

How It Works in Practice

The operational workflow of INTENT can be broken down into four stages:

1. Data acquisition: The vehicle’s perception stack streams synchronized measurements (GPS, IMU, camera‑derived traffic‑light status, and surrounding vehicle trajectories) at a fixed frequency.
2. Pre‑processing & windowing: A rolling buffer collects the last n timesteps (e.g., 20 × 100 ms samples) and normalizes each feature to a common scale.
3. Sequential inference: The buffered sequence is fed into the LSTM. Hidden states evolve as the network processes each timestep, effectively “reading” the driver’s behavior over the past two seconds.
4. Decision output: The final hidden state passes through the classifier, yielding a probability for each maneuver. A thresholding rule selects the most likely intention, which downstream modules (trajectory planner, safety controller) can consume.

What sets INTENT apart from prior work is its disciplined ablation study. The researchers systematically removed or altered each input modality, demonstrating that the full multi‑modal configuration is essential for reaching near‑perfect accuracy. Moreover, the LSTM’s ability to retain information over the entire window eliminates the need for handcrafted temporal features, simplifying the pipeline while boosting robustness.

Evaluation & Results

To validate INTENT, the authors employed the InD (Intersection Drone) dataset, a large‑scale collection of aerial videos annotated with vehicle trajectories and maneuver labels across diverse intersection layouts. The evaluation protocol involved:

• Splitting the dataset into 70 % training, 15 % validation, and 15 % test sets, ensuring geographic and traffic‑pattern diversity.
• Measuring classification accuracy, precision, recall, and F1‑score for each maneuver class.
• Running a comprehensive ablation analysis that toggled input streams (e.g., removing traffic‑light state) and altered LSTM depth.

Key findings include:

• Overall accuracy: 99.71 % on the held‑out test set, surpassing the next best baseline (a CNN‑RNN hybrid) by more than 4 percentage points.
• Class‑wise performance: Precision and recall above 99 % for straight and right‑turn predictions; left‑turn detection, historically the hardest due to tighter turning radii, achieved 99.5 % recall.
• Ablation insights: Removing traffic‑light information dropped accuracy to 96.2 %; excluding surrounding‑vehicle trajectories reduced it further to 93.8 %, confirming the synergistic value of multi‑modal data.

These results demonstrate that INTENT not only excels in controlled benchmark conditions but also maintains resilience when individual sensor streams degrade—a realistic scenario for on‑road deployment.

Why This Matters for AI Systems and Agents

For engineers building autonomous‑driving stacks, INTENT offers a plug‑and‑play intention module that can be integrated upstream of trajectory planners. Early intention signals enable:

• Proactive safety maneuvers: The vehicle can begin deceleration or lane‑change preparations before a turn is executed, reducing collision risk.
• Improved traffic flow: Predictive intent allows smoother coordination with surrounding agents, minimizing unnecessary stops at intersections.
• Robustness to perception failures: By relying on temporal patterns rather than a single frame, the system tolerates brief occlusions or sensor noise.

From a product perspective, integrating INTENT aligns with the broader vision of UBOS platform overview, where modular AI components are orchestrated through a unified workflow engine. Developers can expose the intention predictions as API endpoints, feeding them into downstream services such as Workflow automation studio for scenario‑based testing or real‑time decision logic.

Moreover, the framework’s reliance on standard sensor inputs means it can be combined with OpenAI ChatGPT integration to generate natural‑language explanations of why a particular maneuver was predicted—useful for debugging, compliance reporting, or driver‑assist handover interfaces.

What Comes Next

While INTENT sets a new benchmark, several avenues remain open for extending its impact:

• Multi‑agent intention modeling: Current work focuses on a single ego vehicle. Future research could predict the intentions of surrounding traffic participants simultaneously, enabling cooperative planning.
• Longer prediction horizons: Extending the forecast window beyond two seconds may benefit highway merging scenarios, though it will require richer context modeling.
• Domain adaptation: Training on the InD dataset yields excellent results, yet real‑world deployments must handle diverse weather, lighting, and cultural driving norms. Techniques such as unsupervised domain adaptation or continual learning could bridge this gap.
• Edge‑optimized architectures: Deploying LSTMs on automotive‑grade hardware demands quantization and pruning strategies to meet latency constraints without sacrificing accuracy.

Practitioners interested in experimenting with INTENT can explore the Enterprise AI platform by UBOS, which provides pre‑configured containers for LSTM training, data ingestion pipelines, and model serving. Start‑ups looking for rapid prototyping may find the UBOS for startups offering valuable resources, while SMBs can leverage UBOS solutions for SMBs to integrate intention prediction into existing fleet management tools.

Finally, the authors have made their code and trained weights publicly available alongside the paper. Interested readers can dive deeper into the methodology by consulting the original publication: INTENT paper on arXiv.

Call to Action

Ready to bring cutting‑edge intention prediction into your autonomous‑driving stack? Visit the UBOS homepage to explore our AI solutions, or reach out through the About UBOS page to discuss a custom integration.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.