Unsupervised LiDAR Object Detection via Iterative Self-Labeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
LiDAR systems face challenges in dense sensor costs and sparse point cloud data, especially at far distances, which complicates 3D perception and object detection, particularly in poor weather conditions and with limited training data for machine learning models.
Innovation Solution
An unsupervised iterative training method for a detector model that processes LiDAR point clouds to generate new object positions and bounding boxes, matching them to forecasted positions, updating object tracks, and filtering to create a training set for retraining the model, thereby improving object detection accuracy in sparse regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling is used to generate training data for machine learning models, then the model can learn to group LiDAR points belonging to the same object, but the amount of training data is limited and manual data generation is time-consuming and expensive
Solution Approach 1:
The system performs self-labeling by automatically generating training data through its own detection and tracking capabilities. The detector model processes LiDAR point clouds to generate bounding boxes, the tracker associates these with object tracks, and the system uses this self-generated data for iterative retraining, eliminating dependence on manual labeling
Solution Approach 2:
The system implements iterative retraining where detection results are fed back into the training process. The detector model is retrained using bounding boxes from the current iteration, and this cycle repeats to continuously improve detection accuracy, creating a self-improving feedback loop
2Measurement precision
If dense LiDAR sensors are used to capture high-quality point clouds for better 3D perception, then object detection accuracy improves, but the sensor cost increases significantly
Solution Approach 1:
The system creates virtual dense point clouds by generating synthetic training data that simulates dense LiDAR measurements. Through iterative detection and tracking, the system produces labeled bounding boxes that serve as copies of what dense sensing would provide, enabling training without actual dense sensors
Solution Approach 2:
The system changes the parameter of point cloud density through computational processing rather than physical sensing. By using iterative detection algorithms and tracking over multiple frames, the system effectively increases the density of detectable objects from sparse input data
3Quantity of substance
If low-beam LiDAR is used to reduce sensor cost, then the point clouds captured are sparse especially at far distances, which complicates 3D perception and object detection
Solution Approach 1:
The system maintains continuous tracking of objects across multiple LiDAR frames. By continuously associating detection results with object tracks and iteratively refining bounding boxes through retraining, the system accumulates useful information over time to overcome the sparsity of individual frames
Solution Approach 2:
The system performs preliminary detection and tracking to generate initial bounding boxes and object tracks before using them for iterative retraining. This preliminary action creates a foundation of labeled data that guides subsequent detection improvements in sparse regions
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Enhances object detection accuracy in LiDAR systems by automatically generating training data and iteratively updating the detector model, improving performance in sparse regions and reducing reliance on manual labeling.
Implementation Method 1
LiDAR sensors use time-of-flight to obtain measurements of a surrounding region. Specifically, a LiDAR sensor may scan the environment by rotating emitter-detector pairs (e.g., beams) around the azimuth. At every time step, each emitter emits a light pulse which travels until the beam hits a target, gets reflected, and is received by the detector. Distance is measured by calculating the time of travel.
Data Source
AI summary
Unsupervised object detection from lidar point clouds includes forecasting a set of new positions of a set of objects in a geographic region based on a first set of object tracks to obtain a set of forecasted object positions, and obtaining a new LiDAR point cloud of the geographic region. A detector model processes the new LiDAR point cloud to obtain a new set of bounding boxes around the set of objects detected in the new LiDAR point cloud. Object detection further includes matching the new set of bounding boxes to the set of forecasted object positions to generate a set of matches, updating the first set of object tracks with the new set of bounding boxes according to the set of matches to obtain an updated set of object tracks, and filtering, after updating, the updated set of object tracks to remove object tracks failing to satisfy a track length threshold, to generate a training set of object tracks. The object detection further includes selecting at least a subset of the new set of bounding boxes that are in the training set of object tracks, and retraining the detector model using the at least the subset of the new set of bounding boxes.