Selecting Spots by Explicitly Predicting Intention from Motion History Improves Performance in Autonomous Parking

This paper proposes an autonomous valet parking pipeline that improves task completion, prediction accuracy, and social acceptance by explicitly predicting other agents' parking intentions from their motion history, demonstrating that this approach outperforms methods relying on short-term motion prediction or implicit intention modeling.

The Gist

Imagine you are driving a self-driving car into a busy, crowded parking lot. Your goal is simple: find a spot, get out, and leave. But there's a catch: you have to do this without crashing into other cars, and you need to be polite enough not to cut someone off who was already heading for that same spot.

This paper is about teaching a self-driving car how to be a good driver by learning to read minds (or at least, guess intentions) based on how other cars have moved in the past.

Here is the breakdown of their idea using simple analogies:

The Problem: The "Guessing Game" of Parking

In normal traffic, cars mostly follow lanes. But in a parking lot, it's chaotic. Cars might be backing up, turning sideways, or driving in circles looking for a spot.

• Old Way (The "Crystal Ball" Approach): Previous methods tried to predict where a car will go by looking at where it is right now and guessing its future path. It's like trying to guess where a basketball player will jump just by looking at their feet. If they suddenly change direction, you get it wrong.
• The New Way (The "Detective" Approach): The authors say, "Don't just look at where they are going next; look at where they have been." By analyzing a car's recent history (its "motion history"), the car can figure out the driver's intention (e.g., "Ah, that car is slowly backing up and turning; they are definitely trying to park in that specific spot").

The Core Idea: Reading the "Belief Map"

The self-driving car builds a mental map called a Belief Map. Think of this like a game of Battleship or a heat map on a weather app.

1. The Detective Work: The car watches other vehicles. If it sees a car moving slowly toward a spot, it marks that spot on its map as "Likely to be taken soon."
2. Reconstructing the View: Sometimes, the car can't see everything because of pillars or other cars (occlusion). The authors created a clever trick where the car uses what it does see to guess what it doesn't see. It fills in the blind spots with "probabilities" (e.g., "I can't see that car, but based on where it was 5 seconds ago, it's probably hiding behind that pillar").
3. The Decision: Once the car knows, "Hey, that red car is definitely going to park in Spot A," it avoids Spot A. Instead, it picks Spot B, which is empty and safe.

The "Social" Aspect: Being Polite

The paper emphasizes Social Acceptance.

• The Rude Driver: A car that ignores intentions might race to a spot, forcing the other driver to slam on their brakes or miss their spot. This is "stealing" a spot.
• The Polite Driver: By predicting the other car's intention, our self-driving car waits or chooses a different spot. It lets the other driver have the spot they clearly wanted. This makes the whole parking lot flow smoother and reduces stress.

The Experiment: A Virtual Parking Lot

The researchers built a video game-like simulation to test this.

• They created a parking lot with "reactive" agents (other cars that would brake if the self-driving car got too close).
• They compared their "Mind-Reading" method against two other methods:
  1. The "Future-Guesser": Tries to predict the future path first, then guesses the intention. (Like the authors say, this is circular and often wrong).
  2. The "Black Box": Tries to learn everything at once without explicitly thinking about intentions.

The Result: The "Mind-Reading" method won. It was better at:

• Not crashing.
• Not stealing spots from other drivers.
• Finding a spot faster because it didn't waste time fighting for a spot that was already taken.

The "Secret Sauce": Bezier Curves

To make these predictions smooth, the car uses something called Cubic Bézier curves.

• Analogy: Imagine drawing a line with a flexible ruler. If you just draw a straight line (Constant Velocity), it looks robotic and stiff. If you use a flexible ruler, you can draw the smooth, natural curve a human driver would make when turning into a spot. The authors found this mathematical tool was the best balance between being fast to calculate and looking like a real human driver.

The Bottom Line

The paper concludes that in the messy, unpredictable world of parking, you can't just guess where a car is going next; you have to understand what it wants to do.

By explicitly predicting intentions from the past, a self-driving car can act more like a considerate human driver, making the parking lot safer and less frustrating for everyone. It's the difference between a robot that just follows rules and a driver who understands the "social rules" of the road.

Technical Summary

Here is a detailed technical summary of the paper "Selecting Spots by Explicitly Predicting Intention from Motion History Improves Performance in Autonomous Parking."

1. Problem Statement

The paper addresses the Autonomous Valet Parking (AVP) problem, where an autonomous vehicle (the "ego" agent) must navigate a parking lot, find an available spot, negotiate with other vehicles, and park without human supervision.

Key Challenges:

• Ambiguity of Intent: Unlike highway driving where lanes constrain movement, parking lots allow for diverse, long-term goals (parking in any of many spots). Short-term motion prediction (e.g., where a car is heading in the next 2 seconds) is often insufficient to determine which specific spot a vehicle intends to occupy.
• Lack of Data: There is a scarcity of large-scale, high-quality parking datasets compared to highway driving data, making training robust trajectory predictors difficult.
• Social Interaction: The ego must avoid "stealing" spots from other agents and minimize interruptions to their plans to ensure social acceptability.
• Observation Limitations: Real-world sensors suffer from occlusion, meaning the ego cannot see the full state of the environment or the semantic layout (e.g., which spots are occupied) for other agents.

2. Methodology

The authors propose a modular AVP pipeline that explicitly predicts the intention (target parking spot or exploration point) of other dynamic agents based on their motion history, rather than inferring intention implicitly from future trajectory predictions.

A. Core Pipeline Components

1. BEV Reconstruction with Belief Maps:

   • Since the ego cannot directly observe the full semantic Bird's-Eye View (BEV) of other agents (due to occlusion), the system maintains a probabilistic belief map ($b_t$) representing the likelihood of each parking spot being occupied.
   • The system updates this belief based on current observations (vacant vs. occupied spots) and historical data.
   • Crucially, it synthesizes the missing semantic BEV information for other agents by "filling in" unobserved spots with virtual vehicles if the belief map suggests they are occupied. This allows the use of existing intention models that require full semantic context.

2. Explicit Intention Prediction:

   • The system utilizes a learned Convolutional Neural Network (CNN) model (adapted from [23]) to predict the probability of an agent parking in a specific spot or exploring a specific point.
   • Input: Motion history, candidate spots, distance to spots, heading alignment, and the reconstructed semantic BEV.
   • Output: A probability distribution over candidate parking spots ($\eta_{k,i}$) and exploration points.

3. Intention-Conditioned Trajectory Prediction:

   • Once intentions are predicted, the system forecasts the future trajectories of other agents.
   • It uses cubic B-spline curves (Bézier curves) conditioned on the predicted goal (the target spot center and heading).
   • This approach balances computational efficiency with geometric feasibility, outperforming constant velocity and Hybrid A* baselines in accuracy.

4. Decision Making & Planning:

   • Spot Selection: The ego selects a parking spot that is vacant and has a low probability of being claimed by a dynamic agent (based on the predicted intentions).
   • Planning: If a safe path to a non-contended spot exists, the ego executes a Hybrid A* planner to park.
   • Exploration: If no safe spot is found, the ego explores the lot to find new opportunities.
   • Idling: If blocked, the ego waits.

B. Simulation Environment

To test the method, the authors built a simulation with:

• Reactive Agents: Other vehicles and pedestrians react to the ego's actions using a "passiveness" parameter (how far ahead they look for collisions before braking).
• Occlusion-Aware Sensing: A ray-tracing model simulates the ego's limited field of view and occlusion by other vehicles.
• Realistic Constraints: The ego has imperfect trajectory prediction capabilities and must operate without perfect ground truth of other agents' future paths.

3. Key Contributions

1. Novel BEV Reconstruction: A method to synthesize semantic BEV information for surrounding agents using only ego-centric observations and a belief map, enabling the deployment of learned intention models in realistic, occluded environments.
2. Explicit Intention-Based Spot Selection: A pipeline that explicitly predicts intentions from motion history to avoid spot contention. This outperforms methods that infer intentions implicitly (end-to-end) or derive them from future trajectory predictions.
3. Intention-Conditioned Trajectory Prediction: A trajectory generation method using cubic Bézier curves conditioned on predicted goals, which achieves a superior balance between low computation time, high prediction accuracy, and task completion rates.
4. Comprehensive Evaluation: A simulation framework with reactive agents and occlusion modeling that demonstrates the superiority of explicit intention reasoning in parking scenarios compared to existing baselines.

4. Experimental Results

The method was evaluated against three baselines:

• Explicit, Future ([20]): Infers intention from predicted future trajectories.
• Implicit ([23]): Uses an end-to-end model on the ego only.
• Trajectory Variants: Constant Velocity, Hybrid A*, and Learned Transformer models.

Key Findings:

• Task Completion: The proposed method achieved a significantly higher success rate (approx. +9% for non-reactive agents, +1% for reactive agents) compared to baselines.
• Social Acceptance: The method "stole" fewer spots from other agents (approx. -4%) and interrupted their plans less frequently, indicating more socially acceptable behavior.
• Prediction Accuracy: The intention-conditioned Bézier curves yielded lower minADE (Average Displacement Error) and minFDE (Final Displacement Error) compared to constant velocity and Hybrid A* baselines, and performed comparably to complex learned models.
• Computation: While spot selection took slightly longer due to neural network inference and BEV reconstruction, the total time was negligible compared to the path planning time (Hybrid A*).

5. Significance and Conclusion

The paper argues that in environments with lax driving regulations and diverse goals (like parking), explicit intention prediction is superior to implicit reasoning or short-term motion extrapolation.

• Why it matters: Short-term motion cues (e.g., a car moving forward) are ambiguous; a car could be moving forward to park in the next spot, the one after, or to exit the lot. Explicitly modeling the long-term goal from motion history resolves this ambiguity.
• Broader Impact: The findings suggest that similar explicit intention modeling could benefit other social navigation tasks, such as service robots in indoor spaces or autonomous delivery systems, where agents must negotiate shared spaces with ambiguous long-term goals.

Limitations:

• Current implementation is not yet real-time due to the computational cost of Hybrid A* path planning.
• Experiments were conducted in simulation with predefined agent behaviors rather than real humans.
• The intention model does not yet account for pedestrians, vehicles exiting spots, or non-standard vehicle types (bikes, trucks).