LiDAR Motion Prediction Using Spatiotemporal Point-Cloud Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous driving systems face challenges in accurately predicting motion and estimating environmental states due to dependencies on object detection, which struggles with generalization to unseen classes and discretization errors in occupancy grid maps.
Innovation Solution
A neural network-based system that processes 3D point clouds to estimate motion information by encoding spatiotemporal features without approximating point clouds, using a spatiotemporal pyramid network and novel consistency losses to improve prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If object detection-based state estimation is used, then semantic information can be obtained, but the system fails to generalize to unseen object classes in open-set scenarios
Solution Approach 1:
The patent extracts the motion prediction task from the object detection dependency by using occupancy grid maps to represent environmental states. This separation allows the system to predict motion of all objects including unseen classes without requiring them to be detected and classified first, thereby improving generalization to open-set scenarios while maintaining reliability through the occupancy-based representation
Solution Approach 2:
The patent introduces occupancy grid maps as an intermediary representation between raw sensor data and motion prediction. This intermediary structure enables the system to handle unseen object classes by representing them as occupied grid cells rather than requiring explicit detection and classification, thus resolving the contradiction between adaptability to new classes and detection reliability
2Reliability
If occupancy grid maps are used to represent environmental states, then drivable space can be specified, but correspondence between grid cells across time is difficult to establish
Solution Approach 1:
The patent applies preliminary action by establishing correspondence between occupancy grid cells across time through predicted trajectories. Before performing motion prediction, the system sets up the temporal relationship framework using the occupancy grid structure, which enables consistent tracking of object positions over time and simplifies subsequent motion estimation without requiring complex real-time correspondence matching
3Reliability
If BEV map with fixed grid cells is used, then cell motion and category information can be included, but discretization errors occur where continuous environment is not accurately captured
Solution Approach 1:
The patent applies local quality by allowing different occupancy grid cells to have different properties (motion vectors, categories) while maintaining the overall discrete structure. This enables the system to represent continuous environmental variations through localized cell-specific attributes, improving measurement precision within the discrete grid framework without sacrificing the computational efficiency of the BEV representation
Data Source
AI summary
The present disclosure provides a system and a method for motion prediction for autonomous driving. The system disclosed herein provides an efficient deep-neural-network-based system to jointly perform perception and motion prediction from 3D point clouds. This system is able to take a pair of LiDAR sweeps as input and outputs for each point in the second sweep, both a classification of the point into one of a set of semantic classes, and a motion vector indicating the motion of the point within the world coordinate system. The system includes a spatiotemporal pyramid network, which extracts deep spatial and temporal features in a hierarchical fashion. The training of this system is regularized with spatial and temporal consistency losses. Thus providing an improved motion planner for autonomous driving applications.