Gaussian Process Memory Buffering for High-Rate Multi-Object Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current multi-object tracking techniques, such as re-identification-based and cross-attention-based methods, are inefficient for autonomous driving applications due to high computational costs and memory requirements, limiting their scalability and accuracy in tracking multiple objects at high frame rates.
Innovation Solution
The use of a Gaussian process (GP) based memory buffer update technique that computes a joint Gaussian distribution between current and previous object representations to efficiently remove outdated objects, allowing for flexible memory management and tracking of multiple objects with reduced memory usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If re-identification-based or cross-attention-based methods are used for multi-object tracking, then tracking accuracy is improved, but computational cost and memory requirements increase significantly
Solution Approach 1:
The patent extracts only the essential features needed for tracking (position, velocity, acceleration) from object representations, discarding redundant information. This is achieved by using a simplified state vector that contains only the minimum necessary data for prediction, rather than storing complete object representations used in re-identification methods.
Solution Approach 2:
Instead of matching current detections with stored representations (forward matching), the patent inverts the approach by predicting future positions and matching predictions with actual detections. The Gaussian process predictor generates expected object states, and deviations from these predictions trigger updates, reversing the traditional matching paradigm.
2Measurement precision
If re-identification-based or cross-attention-based methods are used for multi-object tracking, then tracking accuracy is improved, but memory requirements increase significantly
Solution Approach 1:
The patent extracts only the essential features needed for tracking (position, velocity, acceleration) from object representations, discarding redundant information. This is achieved by using a simplified state vector that contains only the minimum necessary data for prediction, rather than storing complete object representations used in re-identification methods.
Solution Approach 2:
The patent changes the parameters stored in memory from detailed object representations (images, extensive feature vectors) to compact state vectors containing only kinematic parameters (position, velocity, acceleration). This parameter transformation dramatically reduces memory requirements while maintaining tracking functionality through physics-based prediction.
3Measurement precision
If traditional multi-object tracking methods are used, then tracking accuracy is improved, but frame rate decreases due to high computational cost
Solution Approach 1:
The patent implements periodic updating where the Gaussian process predictor continuously generates predictions without requiring frequent memory lookups. Updates occur periodically when detections deviate from predictions or when objects enter/exit the scene, rather than requiring processing of every frame, enabling high frame rates while maintaining accuracy.
Solution Approach 2:
The patent performs preliminary prediction using the Gaussian process model to estimate object states before actual detection matching occurs. This preliminary action filters out unnecessary computations by pre-determining expected object positions and characteristics, so that full processing is only needed when predictions and actual detections diverge.
Data Source
AI summary
An apparatus for multi-object tracking determines a current representation of a current object in a current image. The apparatus computes a joint Gaussian distribution between the current representation of the current object and a previous representation stored in one or more memory buffers, wherein the previous representation was determined from a previous image. The apparatus updates the one or more memory buffers based on the joint Gaussian distribution. For example, the apparatus determines whether to remove or replace the previous representation in the one or more memory buffers based on values of a covariance matrix of the joint Gaussian distribution.