TL;DR
This paper introduces a novel method for autonomous vehicles to recognize intersection types using a 2D multi-layer laser scanner. By building a "static local coordinate occupancy grid map" (SLOGM) that accounts for ego-motion and removes dynamic objects, the system achieves high accuracy in classifying six different intersection types, outperforming previous methods.
What Problem Does This Paper Address?
Navigating urban environments safely requires autonomous vehicles to accurately understand their surroundings, especially at intersections. Different intersection types (e.g., merge, diverge, T-junctions) demand distinct driving strategies. Existing methods for intersection recognition often rely on cameras (computationally expensive, sensitive to lighting/weather) or 3D laser scanners (expensive, bulky). While 2D laser scanners offer a more practical solution, previous approaches using them struggled with computational complexity and were limited to simpler intersection scenarios. This paper aims to bridge this gap by proposing an efficient and accurate intersection recognition system using a readily available 2D multi-layer laser scanner.
Methods
The core of the proposed method lies in building a Static Local Coordinate Occupancy Grid Map (SLOGM). This involves two key steps:
- Static Local Coordinate Occupancy Grid Map (SLOGM) Building:
- Local Coordinate System: Instead of a global map, an occupancy grid map is built relative to the ego-vehicle's local coordinate system. This significantly reduces memory requirements and focuses on the immediate environment.
- Dynamic Binary Bayes Filter: A standard binary Bayes filter struggles with local maps because the environment appears to move due to ego-motion. The authors adapt a dynamic binary Bayes filter to handle this, allowing the map to update correctly despite the vehicle's movement.
- Dynamic Object Removal: Moving objects (like other vehicles) can interfere with intersection classification. The paper introduces a method using a "score" based on the intersection's overlap with previous frames to identify and remove dynamic objects, resulting in a "static" map.
- Intersection Classification:
- Feature Extraction: The generated SLOGM serves as the primary feature for classification.
- Nearest Neighbor (NN) Classifier: A Nearest Neighbor classifier is employed.
- New Similarity Measure: A novel similarity measure, SIM(L, M), is developed to compare two SLOGMs. This measure quantifies the overlap of "free" (drivable) space between two maps, with a value of 1 indicating high similarity. This measure is crucial for the NN classifier to effectively distinguish between different intersection types.
Datasets
The paper uses a dataset collected from real-world driving scenarios.
- Data Collection: A vehicle equipped with an IBEO LUX2010 multi-layer laser scanner (4 layers) and a camera was used.
- Dataset Size: 1213 SLOGM samples were collected, covering an area of 80m x 50m with 0.25m x 0.25m cells.
- Ground Truth: A single camera was used to gather ground truth for the intersection types.
- Training and Testing: A five-fold cross-validation approach was used, running 100 times, to evaluate the system's performance.
Key Insights and Ideas
- Local Maps are Key: Shifting from global to local occupancy grids is a critical step for efficient and practical intersection mapping in autonomous vehicles.
- Ego-Motion is a Challenge, Not a Roadblock: The dynamic binary Bayes filter effectively addresses the inherent problem of ego-motion in local map building.
- Static is Smarter: Removing dynamic objects from the map is essential for robust intersection classification, as their presence can obscure crucial structural information.
- Shape Matters: The SLOGM captures the geometric "shape" of the intersection, and the proposed similarity measure effectively leverages these shape differences for classification. T-junctions, with their distinct shapes, are particularly easy to classify.
Key Results
- High Accuracy: The proposed method achieved a high average True Positive Rate (TPR) of 89.15% across all six intersection classes.
- Class Performance: The highest TPR was observed for T-junctions (T1: 91.575%, T2: 91.532%), attributed to their unique SLOGM shapes. Merge-roads (M) and Highways (H) were found to be more challenging due to limited sensor FOV and commonalities in appearance.
- Outperforms Prior Work: The method significantly outperformed previous approaches ([6, 8]) which reported average TPRs of 54.61% and 46.91% respectively, even when those methods used higher-resolution 3D laser scanners. This highlights the effectiveness of the SLOGM and the proposed similarity measure.
Limitations and Future Work
- Sensor Limitations: The paper acknowledges that the limited Field of View (FOV) of the laser scanner can make it difficult to observe the entire intersection, particularly for merge and highway scenarios. This impacts classification accuracy for these types.
- Computational Cost: While the local OGM approach reduces memory, the computational cost of the dynamic binary Bayes filter and the NN classifier might still be a consideration for real-time implementation on resource-constrained platforms.
- Future Work: The authors suggest that further improvements could be made by exploring more advanced classifiers or by integrating additional sensor modalities to overcome the limitations imposed by the laser scanner's FOV.
Summary updated