LiDAR Feature Map Anchor Pruning for Navigation Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current vehicle navigation systems face inefficiencies due to the large number of anchors in LiDAR point cloud feature maps, which increases processing overhead and time, necessitating a method to prune unnecessary anchors.
Innovation Solution
A navigation system that receives a LiDAR point cloud, extracts a feature map with anchor points, smooths the feature map, computes pixel density and values within anchor boxes, and prunes anchors based on predetermined thresholds to reduce the number of anchors before passing the map to object detection processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a large number of anchors are used in the feature map, then object detection coverage is improved, but processing overhead and time increase significantly
Solution Approach 1:
The patent applies preliminary action by performing anchor pruning before the main object detection process. The system pre-processes the feature map to identify and remove low-value anchors based on density thresholds and spatial distribution criteria, thereby reducing the computational burden in subsequent detection stages while preserving anchors that are most likely to contribute to accurate object detection
Solution Approach 2:
The patent extracts and removes unnecessary anchors from the feature map based on specific criteria such as pixel density thresholds and spatial redundancy. By taking out only the essential anchors and eliminating the rest, the system maintains detection coverage for relevant objects while significantly reducing the total number of anchors that require processing
2Reliability
If a large number of anchors are used in the feature map, then object detection coverage is improved, but processing overhead increases
Solution Approach 1:
The system performs preliminary filtering of anchors based on density maps and spatial distribution analysis before the main detection algorithm executes. This pre-processing step simplifies the subsequent processing by reducing the number of anchors that need to be evaluated by complex detection algorithms, thereby reducing overall processing overhead while preserving detection accuracy
Solution Approach 2:
The patent extracts only the necessary anchors from the feature map by applying density thresholds and spatial redundancy criteria. This extraction process removes excessive anchors that would increase processing overhead without contributing meaningfully to detection coverage, thus simplifying the processing burden while maintaining essential detection capabilities
3Productivity
If anchor pruning is performed based on density thresholds, then processing efficiency is improved, but risk of removing relevant anchors increases
Solution Approach 1:
The patent implements feedback mechanisms by using region proposal networks and detection results to guide the anchor pruning process. The system continuously refines the density thresholds and pruning criteria based on feedback from detection performance, ensuring that anchors critical for detecting specific object types are preserved while still removing redundant anchors to maintain processing efficiency
Solution Approach 2:
The patent applies local quality by using different density thresholds and pruning strategies for different spatial regions and object types. Instead of applying a uniform pruning criterion across the entire feature map, the system adapts the pruning parameters locally based on the specific characteristics of each region, thereby preserving relevant anchors in critical areas while aggressively pruning redundant anchors in less important regions
Data Source
AI summary
Embodiments include a method for pruning anchor points from a feature map generated from a Light Detections And Ranging (LiDAR) point cloud, the method comprising: receiving, by a navigation system, a LiDAR point cloud from a LiDAR sensor, the LiDAR point cloud comprising data representing one or more objects in physical surroundings detected by the LiDAR sensor; extracting, by the navigation system, a feature map from the LiDAR point cloud, the feature map comprising a plurality of anchor points, each anchor point defined by an anchor box; smoothing, by the navigation system, the extracted feature map; determining, by the navigation system, density of pixels within the anchor box of each anchor point; and pruning, by the navigation system, anchor points from the feature map based on a plurality of factors related to the determined density of pixels within the box of each anchor point.