Bird's Eye View Flow Estimation for Autonomous Vehicle Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately estimating the motion of surrounding objects due to issues like false negatives, false positives, and occlusions, particularly when using conventional 'detect then track' approaches, which can lead to unsafe maneuvers.
Innovation Solution
An end-to-end deep learning framework for LIDAR-based flow estimation using bird's eye view embeddings, which encodes consecutive point cloud pairs into 2D BeV 'flow' grids, enabling self-supervised learning without ground truth labels, and improves tracking performance of both dynamic and static objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional 'detect then track' approaches are used, then object detection can be performed using 3-D object detection advancements, but tracking accuracy deteriorates due to geometric consistency errors, false negatives, and false positives
Solution Approach 1:
The patent merges detection and tracking into a unified end-to-end deep learning framework that directly estimates object states from point cloud data, eliminating the errors that arise from sequential processing and geometric consistency requirements between separate detection and tracking modules
Solution Approach 2:
The patent replaces the mechanical geometric consistency checks and manual association algorithms with a learned neural network model that directly predicts object states and trajectories, allowing the system to handle occlusions and appearance changes through learned patterns rather than rigid geometric constraints
2Measurement precision
If 3-D scene flow approaches are used, then dense 3-D flow prediction can be computed, but computational complexity increases and the system requires ground truth labels for training
Solution Approach 1:
The patent extracts only the essential 2-D bird's eye view features needed for velocity estimation rather than computing full 3-D scene flow, removing unnecessary computational complexity while retaining the core functionality for motion estimation
Solution Approach 2:
The patent implements self-supervised learning where the system generates its own training labels by leveraging temporal consistency and motion patterns from the point cloud data itself, eliminating the need for external ground truth annotations while maintaining accurate velocity estimation
3Ease of manufacture
If self-supervised learning is implemented, then ground truth labels are not required for training, but the system must learn consistent motion patterns from unlabelled data
Solution Approach 1:
The system generates its own supervision signals by exploiting temporal consistency constraints and motion physics from unlabelled point cloud sequences, automatically creating training labels without human annotation while maintaining accurate velocity estimation through learned motion patterns
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.