Bird's Eye View Velocity Estimation for Autonomous Vehicles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately estimating the motion of surrounding obstacles due to issues like false negatives, false positives, and occlusions, especially when dealing with dynamic scenes and changing viewpoints, which can lead to unsafe maneuvers.
Innovation Solution
An end-to-end deep learning framework for LIDAR-based flow estimation using a bird's eye view (BeV) representation, which aligns consecutive point cloud data sets into the same coordinate frame, encodes them using a pillar feature network, and performs 2D optical flow estimation to determine object velocity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Difficulty of detecting and measuring
If conventional detect then track approach is used with 3-D object detection, then object detection capability is improved, but tracking accuracy deteriorates due to geometric consistency errors and false positives
Solution Approach 1:
The patent transforms the tracking problem from 3D space to 2D bird's eye view space. By projecting 3D point cloud data onto a 2D BeV grid representation, the system simplifies the tracking task while maintaining essential spatial information. This dimensional reduction eliminates depth-related geometric consistency errors and provides a more stable representation for optical flow estimation, directly resolving the contradiction between detection capability and tracking accuracy.
Solution Approach 2:
The patent replaces traditional geometric consistency-based tracking mechanisms with a learning-based optical flow estimation approach. Instead of relying on rigid geometric models and data association algorithms that are sensitive to detection errors, the system uses a neural network trained to estimate motion directly from sequential BeV representations, achieving more robust tracking accuracy.
2Measurement precision
If LIDAR-based flow estimation is performed in 3D point cloud space, then motion estimation capability is improved, but computational complexity and error propagation increase
Solution Approach 1:
The patent reduces computational complexity by transforming the 3D point cloud data into a 2D bird's eye view grid representation. This dimensional reduction decreases the number of points that need to be processed while preserving the essential spatial and motion information needed for accurate flow estimation, thereby resolving the contradiction between motion estimation accuracy and computational complexity.
3Adaptability or versatility
If conventional tracking methods are used with changing viewpoints and occlusions, then system robustness is improved, but tracking consistency deteriorates due to perceptual aliasing and false motion estimates
Solution Approach 1:
By projecting 3D point cloud data onto a 2D bird's eye view grid, the patent eliminates viewpoint changes and occlusions that cause perceptual aliasing in 3D space. The BeV representation provides a top-down perspective that remains consistent regardless of the sensor's viewing angle, thereby maintaining tracking consistency while preserving robustness to environmental variations.
Solution Approach 2:
The patent creates a simplified 2D copy of the 3D scene in bird's eye view space. This copied representation retains the essential spatial relationships and motion patterns needed for tracking while eliminating the complexities of 3D viewpoint transformations and occlusions, thus resolving the contradiction between robustness and tracking consistency.
Data Source
AI summary
Systems and methods determining velocity of an object associated with a three-dimensional (3D) scene may include: a LIDAR system generating two sets of 3D point cloud data of the scene from two consecutive point cloud sweeps; a pillar feature network encoding data of the point cloud data to extract two-dimensional (2D) bird's-eye-view embeddings for each of the point cloud data sets in the form of pseudo images, wherein the 2D bird's-eye-view embeddings for a first of the two point cloud data sets comprises pillar features for the first point cloud data set and the 2D bird's-eye-view embeddings for a second of the two point cloud data sets comprises pillar features for the second point cloud data set; and a feature pyramid network encoding the pillar features and performing a 2D optical flow estimation to estimate the velocity of the object.