Accurate depth perception is critical for autonomous driving, yet current solutions face a trade-off between the high cost of LIDAR and the reliability issues of passive cameras. This paper introduces a framework that utilizes a low-cost, steerable "Triangulation Light Curtain" to actively probe regions of high depth uncertainty. By recursively updating a probabilistic depth volume with targeted laser measurements, the system significantly improves depth accuracy and resolves common camera errors like scale ambiguity at a fraction of the cost of traditional LIDAR systems.
The Challenge of Depth Perception
The industry is currently split between two primary sensing modalities. LIDAR is highly reliable but prohibitively expensive ($25k+) and often limited by fixed scanning patterns. Conversely, RGB cameras are inexpensive and high-resolution but prone to "scale ambiguity"—the inability to distinguish between a small nearby object and a large distant one—as well as issues with oversaturation. While probabilistic deep learning models can quantify a camera's uncertainty, they are typically passive. This research proposes an active sensing loop that uses camera uncertainty to drive a steerable laser sensor, physically "checking" the areas where the camera is most confused.
The Triangulation Light Curtain
The centerpiece of this approach is the Triangulation Light Curtain (LC). Unlike LIDAR, which scans the entire scene indiscriminately, the LC employs a galvomirror to steer a laser line in synchronization with a rolling shutter camera. This creates a virtual "curtain" in 3D space. If an object intersects this curtain, the camera detects a bright reflection; otherwise, it sees nothing.
Methods
The proposed framework operates through a four-stage recursive process:
- Probabilistic Prior: A deep network (based on Neural RGBD) processes RGB images to generate a Depth Probability Volume (DPV)—a 3D grid where each cell represents the probability of an object's presence.
- Curtain Planning: The 3D DPV is collapsed into a 2D Uncertainty Field (UF). An optimization algorithm then determines the optimal path for the laser curtain to maximize information gain by targeting areas with the highest object probability.
- Observation & Update: The LC takes a measurement. A detected reflection increases the probability at that depth, while no reflection "clears" the space. This is managed via a Recursive Bayesian Update.
- Fusion Network: A 3D Convolutional Neural Network (CNN) fuses the original RGB features with the new LC measurements to produce a refined, high-resolution depth map.
Experimental Results
The system was evaluated using the KITTI dataset (with simulated LC data) and a custom real-world dataset collected using a Jeep equipped with stereo cameras, a physical LC device, and a 128-beam LIDAR for ground truth.
- Accuracy Improvement: On the KITTI benchmark, adding LC fusion to a monocular setup reduced the Root Mean Square Error (RMSE) from 3.176m to 2.458m.
- Efficiency: The uncertainty-aware planning reached stable depth estimates in just 5 iterations, compared to 25 for a non-adaptive planar sweep.
- LIDAR Enhancement: The framework can also upsample sparse LIDAR data, improving the RMSE of a low-resolution scan from 1.118m to 0.702m.
- Real-World Robustness: The system effectively corrected scale errors for various objects in outdoor environments at distances up to 15 meters.
Conclusion and Future Work
This work demonstrates that active, adaptive sensors can bridge the gap between passive cameras and expensive LIDAR. By focusing sensing resources on areas of high uncertainty, the Triangulation Light Curtain provides a scalable and cost-effective path toward safer autonomous navigation. Future research will focus on modeling "scene flow" to handle high-speed dynamic environments and predicting surface reflectivities to further refine the laser's observation model.
Summary updated