⚡ electrical engineering

LiDAR-based Dynamic Blockage Prediction: A Data-driven Approach for Learning Interactive Bayesian Models

This paper proposes a data-driven approach using an interactive generalized dynamic Bayesian network (I-GDBN) and an interactive Markov jump particle filter (I-MJPF) to predict future LiDAR sensor blockages and detect abnormalities in autonomous vehicles by modeling interactions between multiple vehicles through a high-level vocabulary.

Original authors: Saleemullah Memon, Ali Krayani, Pamela Zontone, Lucio Marcenaro, David Martin Gomez, Carlo Regazzoni

Published 2026-05-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Saleemullah Memon, Ali Krayani, Pamela Zontone, Lucio Marcenaro, David Martin Gomez, Carlo Regazzoni

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a self-driving car as a person trying to navigate a busy street while wearing a blindfold that occasionally gets covered by a hand. Usually, the car's "eyes" (its LiDAR sensors) scan the world with laser beams to see where other cars and people are. But sometimes, a big bus, a pedestrian, or even bad weather blocks those lasers. When that happens, the car goes blind, which is dangerous.

This paper proposes a clever way to teach the car to predict when it's about to go blind, so it can prepare before the blockage actually happens. Here is how they did it, broken down into simple steps:

1. The "Memory Book" (Learning the Rules)

First, the researchers taught the computer to understand how cars move. They didn't just look at raw laser data; they turned the movement of every car into a simple "story" or a vocabulary.

The Analogy: Imagine watching a dance. Instead of recording every tiny muscle twitch, you just note the steps: "Step left, spin, jump."
The Method: They used a smart tool called GNG (Growing Neural Gas) to group similar movements into "clusters." Each cluster is like a word in a dictionary. A smooth, normal drive is a sentence made of normal words. A car that suddenly disappears behind a bus is a sentence with a "missing word" or a "dummy word."

2. The "Group Chat" (Learning Interactions)

Cars don't move in isolation; they interact. If a bus stops, the cars behind it must slow down.

The Analogy: Think of a group chat. If Person A says "I'm stopping," Person B knows to stop too. The researchers built a system where the "words" (movement patterns) of different cars are combined into a global vocabulary.
The Result: They created a special model called an I-GDBN (Interactive Generalized Dynamic Bayesian Network). This is like a super-smart rulebook that understands not just how one car moves, but how the whole group of cars moves together. It learns the "grammar" of traffic.

3. The "Crystal Ball" (Predicting the Future)

Now, the system is ready to test. When a new car appears on the road, the system has to guess what will happen next.

Choosing the Right Book: The system looks at the car's current position and asks, "Which of our memory books (models) does this car look most like?" It picks the best match (either a "normal driving" book or a "blockage" book).
The Particle Filter: To predict the future, the system uses a method called I-MJPF. Imagine throwing thousands of tiny darts (particles) at a board. Each dart represents a possible future path. Some darts land on "normal driving," others on "getting blocked." The system weighs these darts based on the rules it learned earlier. If the darts start clustering around a "blockage" scenario, the system knows a blockage is coming.

4. The "Self-Awareness" (Knowing What You Don't Know)

The coolest part is that this system is "self-aware."

The Analogy: If you are walking and your vision is blocked, a normal person might panic. A "self-aware" person knows, "Ah, I can't see right now, but I remember where I was last, so I can guess where I'll be."
The Result: Even if the sensors fail or the car disappears from view, the system can say, "I predict a blockage is happening here," and keep tracking the car's likely path using its internal logic, rather than just giving up.

Summary of What They Claim

The Problem: Sensors get blocked, causing cars to lose track of their surroundings.
The Solution: A data-driven system that learns the "vocabulary" of traffic movements and interactions.
The Outcome: The system can predict when a sensor will be blocked before it happens and can detect "abnormalities" (like a car disappearing behind an obstacle) by comparing real-time data against its learned rules.
The Benefit: It makes the system more explainable (we know why it made a prediction) and adaptable, allowing it to handle situations where sensors fail to provide data.

In short, they taught the car to read the "body language" of traffic so well that it can predict when its own view will be blocked, keeping it safe even when it can't "see."

1. Problem Statement

Autonomous vehicles (AVs) and Intelligent Transportation Systems (ITS) rely heavily on sensors like LiDAR for perception, localization, and decision-making. A critical challenge in these systems is sensor blockage or occlusion, where physical objects (e.g., buses, pedestrians, other vehicles) or adverse weather conditions obstruct the sensor's field of view.

Consequences: Blockages lead to data loss, causing failures in tracking, navigation, and obstacle detection.
Limitations of Existing Solutions: Current AI/Deep Learning approaches (e.g., CNNs, RNNs) for blockage prediction often suffer from:
- High computational complexity.
- Requirement for large amounts of labeled data.
- Lack of "self-awareness" and explainability (black-box nature).
- Difficulty in handling interactions between multiple dynamic objects in real-time.

The paper proposes a solution to predict future LiDAR blockages and detect high-level abnormalities before they occur, enabling proactive system responses.

2. Methodology

The authors propose a novel framework based on an Interactive Generalized Dynamic Bayesian Network (I-GDBN) combined with an Interactive Markov Jump Particle Filter (I-MJPF). The methodology operates in two phases: Offline Learning and Online Testing.

A. Data Preprocessing & Tracking

Dataset: Uses the "Deep Sense 6G" dataset (Scenario 32), featuring real-world street-level Vehicle-to-Infrastructure (V2I) data with 3D LiDAR point clouds.
Preprocessing: Raw point clouds are filtered (ground removal, cropping, denoising) to extract dynamic objects.
Tracking: A Joint Probabilistic Data Association Filter (JPDAF) is used to track multiple vehicles, handling the uncertainty of associating observations with objects.
- Normal Tracks: Continuous trajectories without blockage.
- Abnormal Tracks: Trajectories that disappear (due to blockage) and reappear. A reassociation algorithm predicts missing positions for these tracks.

B. Offline Learning Phase

Generalized States (GS) Extraction: An Unmotivated Kalman Filter (UMKF) processes tracked vehicles to generate GS vectors (position and velocity).
GNG Clustering: An unsupervised Growing Neural Gas (GNG) algorithm clusters the GSs for each vehicle.
- This creates a "vocabulary" where each cluster represents a discrete state (node) in a GDBN.
- Transition Matrices (TMs) are learned to model the probability of moving between clusters.
Handling Blockages: For abnormal tracks, dummy clusters are introduced at blockage locations. The TMs are updated based on reassociation probabilities to create specific "blockage vocabularies."
Interaction Modeling (I-GDBN Construction):
- A higher abstraction level is created to model interactions between multiple vehicles.
- Cluster Coupling: Discrete states ( $S_t$ ) from individual vehicles are combined to form global "words" ( $G_t$ ).
- A Global Vocabulary and Global Transition Matrix are generated, representing the joint state of the environment (e.g., Vehicle A in state 2 AND Vehicle B in state 3 = Word 10).

C. Online Testing Phase

Generative Model Selection: For a new testing track, the system calculates the Euclidean distance between the current 3D positions and the cluster means of all trained GDBN models (both normal and blockage models). The model with the minimum error ( $m^*$ ) is selected.
Inference via I-MJPF: An Interactive Markov Jump Particle Filter is employed to infer future states:
- Particles represent hypotheses of the system state.
- The filter samples words from the probability distribution of the selected I-GDBN.
- It predicts future positions and detects abnormalities by comparing predicted "words" against observed data.

3. Key Contributions

I-GDBN Framework: Proposes a probabilistic, data-driven model that learns interactions between multiple vehicles, moving from low-level observations to high-level abstractions.
Self-Aware Vocabulary: Introduces a mechanism to modify the model's vocabulary by incorporating "dummy nodes" for blockage scenarios, allowing the system to learn and predict occlusions.
Optimal Model Selection: Implements a dynamic selection strategy during testing to choose the best generative model (normal vs. blockage) based on real-time error minimization, improving prediction accuracy.
Explainability: Unlike black-box deep learning models, the I-GDBN provides an interpretable structure where transitions and uncertainties are explicitly modeled, enhancing self-awareness.

4. Simulation Results

The system was evaluated using the Deep Sense 6G dataset:

Normal Scenarios: The model successfully selected the best generative model (e.g., ID-51 for single tracks, IDs 267/321 for interactive tracks), resulting in predicted words that closely matched estimated words with minimal abnormalities.
Blockage Scenarios:
- When a blockage occurred at a known location, the system selected a pre-trained blockage model (ID-1000) and accurately predicted the blockage with low error.
- Limitation: In "new" situations where the blockage duration exceeded the training data (e.g., frames 18–22 in the test), the system detected abnormalities because the specific duration was not fully captured in the model, though it still utilized the correct blockage model ID.
Metric: Performance was measured using Hamming distance (binary 0 for match, 1 for mismatch) between predicted and estimated words.

5. Significance and Future Work

Safety & Proactivity: The approach enables AVs to anticipate sensor failures rather than just reacting to them, allowing for proactive decision-making (e.g., slowing down or changing lanes before a sensor is fully occluded).
Explainable AI (XAI): By using probabilistic graphical models, the system offers transparency in why a blockage was predicted, which is crucial for safety-critical autonomous systems.
Future Directions: The authors suggest extending the framework to Integrated Sensing and Communication (ISAC) systems by fusing 3D LiDAR data with communication signals for enhanced environmental awareness.

In summary, this paper presents a robust, interpretable, and adaptive framework for predicting LiDAR blockages in multi-vehicle environments, bridging the gap between low-level sensor data and high-level situational awareness.