Sparse Tensor Object Tracking for Memory-Limited Perception

Solution Overview

Problem

Conventional tensor-based object detection systems in computer-implemented perception systems consume significant memory due to the computation of non-zero values for all bins, limiting resolution and extent, which is problematic in computing environments with finite memory resources.

Innovation Solution

A computer-implemented system trained using the Gumbel-softmax algorithm sparsely computes non-zero values for tensor bins, reducing memory consumption by computing zero values for bins with low object likelihood, thereby allowing higher resolution without increasing memory usage.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional tensor computation methods are used to compute values for all bins, then complete object detection coverage is achieved, but memory consumption increases significantly

Engineering Contradiction:

Improveobject detection coverageVSAvoidmemory consumption

Core Design Contradiction:

ReliabilityVSQuantity of substance

Solution Approach 1:

The patent extracts only the essential computational information by identifying and computing values only for bins that contain objects, rather than computing values for all bins. This is achieved through iterative computation where bins are processed and marked as containing objects, and subsequent computations skip bins already identified as containing objects, thereby extracting only the necessary data points needed for object detection while discarding redundant computations.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies partial action by computing tensor values for only a subset of bins rather than all bins. The system performs computations iteratively, processing bins until a predetermined number of bins containing objects are identified, or until all bins have been processed. This partial computation approach reduces memory consumption while maintaining sufficient object detection coverage.

Inventive Principle:

Principle #16Partial or excessive action

2Measurement precision

If tensor resolution is increased to improve detection precision, then object detection precision improves, but memory requirements increase

Engineering Contradiction:

Improveobject detection precisionVSAvoidmemory requirements

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent applies partial action by computing tensor values for only a subset of bins rather than all bins. The system performs computations iteratively, processing bins until a predetermined number of bins containing objects are identified, or until all bins have been processed. This partial computation approach reduces memory consumption while maintaining sufficient object detection coverage.

Inventive Principle:

Principle #16Partial or excessive action

Solution Approach 2:

The patent applies local quality by differentiating the treatment of different bins based on their content. Bins that contain objects receive full computational processing and are marked accordingly, while bins that do not contain objects are skipped in subsequent iterations. This localized approach ensures that computational resources are concentrated on regions of interest (where objects are present) rather than uniformly processing the entire tensor space.

Inventive Principle:

Principle #3Local quality

3Area of stationary object

If tensor extent is increased to cover larger scene areas, then scene coverage is improved, but the number of bins and memory usage increase

Engineering Contradiction:

Improvescene coverageVSAvoidnumber of bins

Core Design Contradiction:

Area of stationary objectVSQuantity of substance

Solution Approach 1:

The patent extracts only the necessary computational work by identifying and processing only those bins within the large tensor that actually contain objects. Even though the tensor covers a large scene area with many bins, the iterative computation process extracts and processes only the subset of bins that are relevant for object detection, thereby managing memory usage efficiently despite the large scene coverage.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies partial action by computing tensor values for only a subset of bins rather than all bins. The system performs computations iteratively, processing bins until a predetermined number of bins containing objects are identified, or until all bins have been processed. This partial computation approach reduces memory consumption while maintaining sufficient object detection coverage.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS12482273B2Detecting, identifying, and tracking objects based upon a tensor with sparsely computed non-zero values

Publication Date: 2025.11.25 GM CRUISE HOLDINGS LLC

AI summary

Technologies for detecting, identifying, and tracking objects in a scene based upon images generated by a camera are described. A computing system obtains a perspective image from the camera and generates a feature image based upon the perspective image. The feature image has X×Y pixels, and each pixel in the X×Y pixels has a feature vector that includes values assigned thereto. The feature image is provided as input to a machine learning model, where the machine learning model generates a tensor based upon the feature image, where the tensor includes several bins that include values that are representative of likelihoods that an object exists in the scene at distances from the camera that correspond to the bins. The machine learning model is trained using a Gumbel-softmax algorithm. The computing system generates the image in the overhead view based upon the tensor output by the machine learning model.