LiDAR Point Cloud Restitching for Discontinuity-Free Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
LiDAR sensors face distortions and artifacts due to scanning discontinuity when aggregating 360-degree scans, leading to incomplete or duplicate object detection, especially with moving objects, and timing mismatches affecting fusion with other sensors.
Innovation Solution
The method involves dynamically restitching LiDAR scans to shift the scanning discontinuity to an uncritical location, allowing object detection only in the unaffected half of the restitched scanning samples, and assigning common timestamps to detections within each quadrant to mitigate timing mismatches.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If a full 360-degree LiDAR scan is aggregated into a single 3D point cloud, then complete environmental coverage is achieved, but scanning discontinuity causes distortions and artifacts at the boundary between start and end of scan
Solution Approach 1:
The patent divides the 360-degree scan into multiple overlapping segments (e.g., front, rear, left, right views). Each segment is processed separately to create multiple 3D point clouds, avoiding the discontinuity issue that occurs when aggregating a full 360-degree scan into a single point cloud.
Solution Approach 2:
The patent transitions from processing a single aggregated 3D point cloud to processing multiple 2D images that are then individually converted to 3D point clouds. This dimensional approach allows each view to be processed independently, eliminating the scanning discontinuity boundary artifacts.
2Quantity of substance
If object detection is performed on the entire aggregated 3D point cloud, then all objects in the scene can be detected, but objects near the scanning discontinuity may be duplicated or missed
Solution Approach 1:
The patent performs object detection on multiple separate 3D point clouds derived from different view segments rather than on a single aggregated point cloud. This segmentation ensures that objects near the scanning discontinuity boundary are not duplicated or missed, as each segment is processed independently.
Solution Approach 2:
The patent creates multiple copies of the scene from different viewpoints (front, rear, left, right), processes each copy separately for object detection, and then fuses the results. This copying approach ensures comprehensive object detection while avoiding discontinuity-related errors.
3Productivity
If LiDAR scans are aggregated at 10 Hz with 100 ms delay between start and end measurements, then real-time processing is enabled, but fast moving objects suffer from significant deformation and duplication
Solution Approach 1:
The patent performs motion compensation for the ego-vehicle before aggregating the scan segments into 3D point clouds. By preliminarily correcting for ego-vehicle movement, the system maintains object position accuracy even when processing scans at 10 Hz with inherent time delays.
Solution Approach 2:
The patent dynamically adjusts the processing approach by treating each view segment independently with its own timing reference. This dynamic handling of temporal information allows the system to maintain accuracy for fast-moving objects while preserving real-time processing capabilities.
4Ease of operation
If detections from different scanning angles are treated as simultaneous, then processing simplicity is maintained, but timing mismatches occur when fusing with other sensors
Solution Approach 1:
The patent assigns accurate timestamps to each 3D point cloud and object detection result based on the actual acquisition time of the corresponding view segment. This preliminary timing assignment ensures that when detections are fused with other sensors, the temporal alignment is accurate without requiring complex post-processing adjustments.
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
This approach effectively reduces distortions and artifacts, improves object detection accuracy by avoiding scanning discontinuity issues, and ensures coherent time stamps for sensor fusion, enabling effective use of existing LiDAR and computer vision algorithms.
Implementation Method 1
The ranging sensors are used to build a map of their surrounding by measuring the time required for a laser beam to be reflected from surrounding objects
Implementation Method 2
Most LiDAR sensors work by horizontally spinning a top-mounted array of lasers, for repeatedly constructing a 3D point cloud of the reflected points
Data Source
Figure 1
Figure 2~3
Figure 4~5
AI summary
The present application relates to the field of range sensors, and more particular to a method, a system and a computer-readable storage medium for object detection in a 3D point cloud representing a scanned surrounding of a vehicle. One aspect relates to a method for object detection in a three-dimensional, 3D, point cloud representing a scanned surrounding of a vehicle. The method comprises determining a first plurality of scanning samples (412) representing the scanned surrounding, the first plurality of scanning samples consisting of a first half of scanning samples and a second half of scanning samples. The method further comprises populating the 3D point cloud with the first plurality of scanning samples. The method further comprises detecting objects in the 3D point cloud contained in one or more of the second half of scanning samples. Preprocessing LiDAR data enables the usage of software tools for processing LiDAR data which require large-scale point clouds, while avoiding the scanning discontinuity (416) that arises every time the output of a scanning lidar is aggregated into a 360 degree scan. Aggregated LiDAR data is subdivided into local sub processing portions. LiDAR scans are repeatedly dynamically "restitched" to move the scanning discontinuity (416) within a LiDAR scan to a location uncritical for the current sub processing step. The proposed approach avoids issues resulting from aggregation-related artifacts in automatic tools such as object detectors, as well as in manual processes such as annotating data, without having to change those tools themselves.