Simulated Velocity Grid Training for Curved Object Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating velocity grids for autonomous vehicles are not precise, leading to inaccurate object tracking, particularly in complex scenarios where objects move on curved trajectories.
Innovation Solution
A computer-implemented method is developed to generate more accurate velocity grids through simulation, where a simulated scene with moving objects is created, and machine-learning algorithms predict and adjust velocities based on differences between predicted and simulated velocities, improving object tracking accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual velocity grid generation is used, then the process is simple and quick, but the precision and accuracy of velocity estimates are insufficient
Solution Approach 1:
The patent creates a simulated copy of the real-world scene including simulated objects, simulated sensor systems, and simulated environments. This simulated velocity grid is generated from the simulated scene and used to train machine learning algorithms, providing precise velocity estimates without requiring complex manual annotation of real data
Solution Approach 2:
The patent replaces the manual mechanical process of creating velocity grids with automated computational processes. Machine learning algorithms automatically generate velocity grids from simulated sensor data, substituting manual human effort with automated systems that provide higher precision and consistency
2Reliability
If manual velocity grid creation is used, then the method is straightforward, but object tracking accuracy is limited
Solution Approach 1:
The patent performs preliminary actions by pre-generating simulated velocity grids from simulated scenes before actual object tracking occurs. The machine learning algorithms are trained in advance on this simulated data, so when deployed for real object tracking, they already possess the learned patterns and achieve higher accuracy without requiring complex real-time processing
Solution Approach 2:
The patent introduces a simulated scene and simulated velocity grid as intermediary elements between real sensor data and object tracking. The simulated environment acts as a mediator that translates real-world scenarios into training data with known ground truth velocities, enabling more reliable tracking without directly complexifying the tracking system itself
3Measurement precision
If simulated velocity grid training is implemented, then object tracking accuracy improves significantly, but the system complexity and computational requirements increase
Solution Approach 1:
The simulated scene generation system serves multiple functions: it creates training data for velocity prediction, generates diverse scenarios including curved trajectories, provides ground truth velocity labels, and can be reused for different machine learning models. This multi-functionality reduces overall system complexity despite the added simulation capability
Data Source
AI summary
A computer-implemented method is provided for training a machine-learning (ML) algorithm that contributes to piloting an autonomous vehicle using a velocity grid generated from a simulation. The method may include simulating a scene. The scene includes a simulated autonomous vehicle and at least one simulated object moving in the scene. The method may also include predicting a velocity of at least one point on the simulated object as it moves in the simulated scene using a ML algorithm. The method may also include comparing the predicted velocity of the at least one point on the simulated object with velocities in the velocity grid generated from a simulation. The method may further include adjusting the ML algorithm to more accurately predict velocity of at least one point on the simulated object as it moves in the scene based on a difference in the predicted velocity compared to the velocity grid, which yields a trained ML algorithm.