LiDAR Slice Loss Training for Vehicle Depth Map Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous driving vehicles face errors in depth data generated by LiDAR systems due to transparency of glass surfaces, leading to incorrect distance calculations and potential unsafe vehicle reactions.
Innovation Solution
A machine learning model is trained with a slice loss function to correct erroneous depth data points, using LiDAR data from point cloud points with depths greater than the average plus a threshold, to improve the accuracy of depth maps and vehicle operation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If LiDAR system is used to generate depth data, then distance measurement capability is improved, but measurement precision deteriorates due to glass transparency causing erroneous depth slices
Solution Approach 1:
The patent applies preliminary action by identifying and masking erroneous depth points before final depth map generation. The system detects points with depths greater than average plus threshold, marks them as erroneous, and excludes them from influencing the final depth estimation, preventing propagation of LiDAR slicing errors through the pipeline
Solution Approach 2:
The patent introduces an intermediary mechanism between raw LiDAR depth data and the final depth map. A machine learning model acts as a mediator that processes depth estimates, identifies erroneous slices using the threshold criterion, and generates corrected depth maps by blending LiDAR data with monocular depth estimates where LiDAR data is erroneous
2Manufacturing precision
If machine learning model is trained with all training point cloud data, then training completeness is improved, but training accuracy deteriorates due to inclusion of erroneous depth points
Solution Approach 1:
The patent applies local quality by differentiating between erroneous and valid training points rather than uniformly processing all data. The system calculates average depth per training image, identifies local outliers exceeding the threshold, and applies different processing paths: erroneous points are masked while valid points are used for supervised training, creating quality-aware training data processing
Solution Approach 2:
The patent uses partial action by selectively applying the slice loss function only to identified erroneous depth regions rather than uniformly across the entire training dataset. This targeted approach focuses computational resources on correcting specific problematic areas while maintaining efficiency
3Measurement precision
If depth map is generated using conventional methods, then processing speed is improved, but measurement precision deteriorates due to incorrect depth slices affecting vehicle distance calculation
Solution Approach 1:
The system performs preliminary identification and masking of erroneous depth points before final depth map composition. By pre-processing the LiDAR depth data to flag suspicious points (depth > average + threshold), the system prevents these erroneous values from contaminating the final output, ensuring accuracy without requiring complete reprocessing
Solution Approach 2:
The patent implements feedback mechanisms where the system continuously monitors depth estimates, compares them against statistical thresholds, and adjusts the depth map generation accordingly. The machine learning model receives feedback from the slice detection process and iteratively refines depth predictions, creating a self-correcting pipeline that maintains accuracy
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
The solution enhances the accuracy of depth map generation and vehicle control by reducing errors in depth data, enabling safer and more efficient navigation around neighboring vehicles.
Implementation Method 1
An autonomous driving vehicle may include a LiDAR system or other sensor system for sensing point cloud data indicative of the existence and location of other objects around the autonomous driving vehicle
Data Source
AI summary
Example systems and techniques are described for controlling operation of a vehicle and training a machine learning model for controlling operation of a vehicle. A system includes memory configured to store point cloud data associated with the vehicle and one or more processors communicatively coupled to the memory. The one or more processors are configured to determine a depth map indicative of distance of one or more objects to the vehicle and control operation of a vehicle based on the depth map. The depth map is based on executing a machine learning model, the machine learning model being trained with a slice loss function determined from training point cloud data having a respective depth that is greater than the average depth for a set of points of the point cloud data plus a threshold.