Laser Point Cloud Gap Filling for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately detecting and recognizing objects in their environment due to gaps in laser data, which can hinder safe navigation and obstacle avoidance.
Innovation Solution
A method and system that generate a two-dimensional range image from laser data points, filling in gaps with values from neighboring pixels and determining normal vectors of object surfaces to provide comprehensive object recognition information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Difficulty of detecting and measuring
If laser data is used for object detection, then object detection capability is improved, but gaps in laser data reduce detection accuracy
Solution Approach 1:
The patent introduces an intermediary process that fills gaps in laser data using information from neighboring pixels and multiple laser scans. This mediator recovers missing data points by interpolating from surrounding valid pixels, thereby maintaining detection capability while improving accuracy in regions where direct laser measurements are absent
Solution Approach 2:
The patent performs preliminary gap-filling and data completion before final object detection and recognition. By pre-processing the laser data to recover missing points using neighboring pixel information and historical scan data, the system ensures that subsequent detection algorithms work with complete datasets, improving overall accuracy
2Loss of information
If normal vectors are determined for object surfaces, then object recognition information is improved, but computational complexity increases
Solution Approach 1:
The patent segments the laser point cloud data into multiple regions or groups before calculating normal vectors. By dividing the data into manageable segments and processing normals locally for each segment, the system reduces overall computational complexity while still capturing comprehensive surface information for accurate object recognition
Solution Approach 2:
The patent calculates normal vectors selectively for critical regions or uses approximate normal calculations where full precision is not required. This partial action approach computes only the essential normal information needed for recognition, reducing computational burden while maintaining sufficient recognition accuracy
3Productivity
If two-dimensional range image is generated from laser data points, then processing efficiency is improved, but information completeness deteriorates due to gaps
Solution Approach 1:
The patent introduces a gap-filling intermediary step that operates on the two-dimensional range image after generation. This mediator identifies missing pixels and fills them using interpolation from neighboring pixels, preserving the processing efficiency of 2D image representation while restoring information completeness by recovering missing data points
Solution Approach 2:
The patent creates a completed copy of the range image that fills in missing information based on patterns from valid regions. By generating this enhanced copy with interpolated pixel values, the system maintains the efficiency of 2D processing while ensuring the copied image contains complete information for accurate object detection
Data Source
AI summary
Methods and systems for object detection using laser point clouds are described herein. In an example implementation, a computing device may receive laser data indicative of a vehicle's environment from a sensor and generate a two dimensional (2D) range image that includes pixels indicative of respective positions of objects in the environment based on the laser data. The computing device may modify the 2D range image to provide values to given pixels that map to portions of objects in the environment lacking laser data, which may involve providing values to the given pixels based on the average value of neighboring pixels positioned by the given pixels. Additionally, the computing device may determine normal vectors of sets of pixels that correspond to surfaces of objects in the environment based on the modified 2D range image and may use the normal vectors to provide object recognition information to systems of the vehicle.