3D Object Detection Using Feature Grid Voting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current 3D object detection methods in laser-based systems face computational challenges due to the sliding window approach, which is deemed intractable in 3D environments, and often rely on 2D representations, failing to fully utilize the sparsity of 3D data.
Innovation Solution
A method that converts 3D point clouds into feature grids, allowing for efficient detection by discretizing space, mapping points to feature vectors, and calculating detection scores through voting, eliminating the need for 2D projections and reducing computational intensity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a sliding window approach is applied in 3D point cloud data, then object detection can be performed in three-dimensional space, but the computational burden becomes intractable due to the additional dimension
Solution Approach 1:
The patent transforms the 3D point cloud data into a 2D depth image by projecting laser points onto an image plane. This dimensionality reduction allows the application of efficient 2D sliding window detection algorithms while still maintaining 3D detection capability. The depth information is preserved in the Z-channel of the depth image, enabling 3D object detection without the computational complexity of direct 3D sliding window approaches.
2Productivity
If 2D projections are used for 3D object detection, then computational intensity is reduced, but the sparsity of 3D data is not fully utilized
Solution Approach 1:
The patent introduces a depth image as an intermediary representation that bridges 3D point cloud data and 2D image processing. The depth image preserves the full 3D spatial information including sparsity patterns, while enabling efficient 2D processing algorithms. By working in the depth image domain, the system maintains 3D data characteristics while leveraging computationally efficient 2D convolution and sliding window operations.
3Measurement precision
If 3D point clouds are directly processed without discretization, then measurement precision is maintained, but processing time increases significantly
Solution Approach 1:
The patent discretizes the continuous 3D point cloud space by projecting it onto a 2D grid structure (depth image). This segmentation of space into discrete pixels with depth values maintains measurement precision while enabling efficient grid-based processing. The discretization allows the use of fast 2D array operations and convolution algorithms, dramatically reducing processing time compared to continuous 3D point cloud manipulation.
Data Source
AI summary
A method and system for detecting objects within a three-dimensional (3D) environment, comprising obtaining a 3D point-cloud representation of the environment, the point-cloud comprising a set of point locations, and converting the point-cloud to a 3D feature grid, wherein converting the point-cloud to the 3D feature grid comprises discretizing the 3D space into cells according to a grid size; determining which cells contain points from the 3D point-cloud; and mapping any that do to a feature vector; and mapping any that do not to a zero feature vector. The method further comprises generating a detection window sized to contain an integral number of the cells and having a set of positions which can be occupied; calculating a detection score for each of the positions by: casting a vote for each cell within the detection window into which at least one point falls, wherein each vote is calculated using the feature vector of the cell and a weighting vector; and summing the votes, and determining whether each position contains an object of interest based on the detection score, each detection window with a detection score greater than a threshold being classified as containing an object of interest.