Compound Radar Measurement Model for Extended Object Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automotive radar measurement models fail to accurately capture both kinematic and extended states of objects due to complex real-world reflections, leading to inaccurate object tracking in autonomous vehicles.
Innovation Solution
A compound measurement model is developed, combining principles of contour and surface models, which includes multiple probabilistic distributions constrained to the object's contour, learned offline and refined online using Bayesian smoothing and online adaptation to improve tracking accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a fixed set of points model is used for extended object tracking, then the object state can be augmented to include both kinematic and extended states, but the data association between fixed points and radar detections becomes complex and non-scalable
Solution Approach 1:
The object is divided into multiple fixed points distributed on its surface, with each point independently generating radar measurements. This segmentation allows the complex object to be modeled through simpler point components, improving tracking precision while managing complexity through modular point-based representation
Solution Approach 2:
The fixed points serve multiple functions: they define the object's spatial extent and orientation, generate virtual radar measurements for comparison with actual detections, and provide a framework for data association. This multi-functionality resolves the contradiction by making the same structural elements serve both tracking precision and complexity management
2Measurement precision
If a contour model is used to reflect measurement distribution along the object contour, then the extended state can be captured, but the model requires much more degrees of freedom and becomes computationally complex
Solution Approach 1:
Different regions of the object contour are assigned different properties based on their local characteristics. Points on the contour have localized measurement generation capabilities tailored to their specific position and orientation, allowing accurate measurement distribution modeling without requiring a uniformly complex model across the entire object
Solution Approach 2:
The contour model is made dynamic by allowing the fixed points and their associated measurement distributions to adapt as the object moves and rotates. This dynamic adjustment reduces the effective degrees of freedom needed at any given time while maintaining accurate measurement distribution representation
3Device complexity
If a surface model is used assuming radar measurements are generated from the inner surface, then computational complexity is reduced, but the model does not capture real-world automotive radar measurements that reflect from the outer surface
Solution Approach 1:
Instead of directly modeling the complex outer surface reflections, the patent creates a simplified copy representation using fixed points on the contour. These points generate virtual measurements that approximate the effect of outer surface reflections, reducing computational complexity while maintaining measurement accuracy through probabilistic measurement models
4Adaptability or versatility
If offline learned compound measurement model is used, then the model can be refined for specific objects, but the model parameters need to be updated online to adapt to new objects
Solution Approach 1:
The system implements feedback loops where actual radar measurements are continuously compared with virtual measurements from the compound model. The discrepancies provide feedback for updating model parameters online, enabling adaptation to new objects while managing complexity through incremental parameter adjustment rather than complete model reconfiguration
Data Source
AI summary
A tracking system for tracking an expanded state of an object is provided. The tracking system executes, for a predetermined time period, a probabilistic filter that iteratively tracks a belief on the expanded state of the object, wherein the belief is predicted using a motion model of the object and is further updated using a compound measurement model of the object. After the predetermined time period, the updated beliefs are smoothed to generate a state-decoupled online batch of training data. The compound measurement model includes multiple probabilistic distributions constrained to lie around a contour of the object with a predetermined relative geometrical mapping to the center of the object. The compound measurement model is updated using the online batch of training data. Further, the tracking system tracks the expanded state of the object based on the updated compound measurement model.