Multi-Object Trajectory Prediction With Selective 3D Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional trajectory prediction systems face instability due to reliance on 2D information, leading to misjudgments and increased computational burden with the introduction of 3D information, and struggle with accurate prediction for multiple target objects.
Innovation Solution
A trajectory predicting method using a generator network with an encoder and decoder architecture, incorporating feature extraction, pooling, and adaptive voxel size determination for 3D point cloud data to process multiple target objects, enhancing trajectory prediction efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If 3D information is introduced to improve detection stability, then detection accuracy is improved, but system computational burden increases causing inability to meet real-time requirements
Solution Approach 1:
The patent extracts only the essential 3D trajectory information needed for prediction while discarding redundant data. The system processes point cloud data to extract key motion characteristics and spatial relationships, maintaining detection stability while reducing computational burden by focusing only on relevant features rather than processing all 3D information.
Solution Approach 2:
The patent applies different processing strategies to different parts of the data. Important trajectory points and critical spatial relationships receive detailed processing, while less critical areas use simplified processing. This localized approach maintains accuracy for key predictions while reducing overall computational load.
2Productivity
If conventional 2D information is used for trajectory prediction, then computational burden is reduced, but prediction accuracy deteriorates due to misjudgments and instability
Solution Approach 1:
The patent transitions from 2D image information to 3D point cloud trajectory information, adding the depth dimension to trajectory data. This dimensional enhancement provides more accurate spatial and temporal characteristics of moving objects, improving prediction accuracy while maintaining computational efficiency through selective feature extraction from the 3D data.
3Adaptability or versatility
If trajectory prediction is performed on multiple target objects, then application coverage is improved, but prediction accuracy deteriorates due to inability to handle multiple objects simultaneously
Solution Approach 1:
The patent segments the trajectory prediction process into independent processing streams for different target objects. Each object's trajectory is extracted and processed separately through the prediction model, allowing accurate prediction for multiple objects simultaneously. The system identifies and tracks each object's point cloud trajectory independently, maintaining prediction accuracy across multiple targets.
Data Source
AI summary
A trajectory predicting method and a computing system for trajectory prediction are provided. In the method, feature extraction is respectively performed on past trajectories of multiple target objects through an encoder to generate first trajectory information of the target objects. A pooling process is performed on the first trajectory information of the target objects to generate second trajectory information of the target objects. The second trajectory information of each target object includes location relationships relative to other target objects. Third trajectory information is obtained from the past trajectories of the target objects. The third trajectory information includes a moving direction, scene information, and/or a moving mode. The predicted trajectories of the target objects are generated according to the second trajectory information and the third trajectory information through a decoder. Accordingly, the accuracy of prediction can be improved.