Object Detection Feature Plane Summation for Embedded Systems

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing object detection systems face challenges in achieving real-time performance due to high computation and memory bandwidth requirements, particularly in embedded systems, when using histogram of oriented gradients (HOG) based techniques for detecting pedestrians and other objects.

Innovation Solution

The system computes feature planes and summations for each image pixel, defining cells with integral multiples of step sizes, and determines feature vectors by summing pixel features, comparing them with classifiers to detect objects, thereby bypassing complex orientation calculations and optimizing memory usage.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If HOG based object detection techniques are used, then object detection accuracy is improved, but computation time and memory bandwidth requirements increase

Engineering Contradiction:
Improveobject detection accuracyVSAvoidcomputation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The image is divided into multiple cells, and each cell is processed independently to compute feature plane summations. This segmentation allows parallel processing of different regions, reducing overall computation time while maintaining detection accuracy through localized feature analysis.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Feature planes are pre-computed for all image pixels before cell formation and classification. By computing gradients, magnitudes, and orientations in advance and storing them in optimized data structures, the system eliminates redundant calculations during the detection phase, significantly reducing real-time computation requirements.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If HOG based object detection techniques are used, then object detection accuracy is improved, but memory bandwidth usage increases

Engineering Contradiction:
Improveobject detection accuracyVSAvoidmemory bandwidth usage
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

Multiple feature planes (gradient magnitude, gradient orientation, and HOG bin assignments) are merged into a unified data structure where feature plane summations are computed and stored together for each cell. This consolidation reduces the number of separate memory accesses and operations, decreasing overall memory bandwidth consumption.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent transforms the traditional HOG computation from a pixel-by-pixel approach into a cell-based approach with feature plane summations. By adding the dimension of cell aggregation and using integral-like images for rapid summation, the system reduces memory access patterns from O(N) to O(1) for many queries, significantly lowering bandwidth requirements.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Measurement precision

If gradient orientation computation is performed using two look-up tables, then computation accuracy is improved, but device complexity and computation time increase

Engineering Contradiction:
Improvegradient orientation accuracyVSAvoidprocessor complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts only the essential gradient information (magnitude and orientation bins) needed for HOG computation, removing the need for complex look-up tables. By directly computing gradient orientations using simplified formulas and rounding to nearest bin centers, the system achieves sufficient accuracy for object detection while dramatically reducing processor complexity and memory requirements for lookup structures.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3224764B1Systems and methods for object detection
Publication Date: 2021.03.03 TEXAS INSTRUMENTS INC
  • EP3224764B1 patent drawingFigure 1~2
  • EP3224764B1 patent drawingFigure 3A~3B
  • EP3224764B1 patent drawingFigure 4~5B

AI summary

In described examples, a method of detecting an object in an image includes computing (802) one or more feature planes of one or more types for each image pixel of the image. Cells (804) are defined in the image, where each cell includes first through nth number of pixels. Starting locations of each cell in the image in horizontal and vertical directions are integral multiples of predefined horizontal and vertical step sizes, respectively. One or more feature plane summations of one or more types are computed (806) for each cell. A feature vector is determined (808) for an image portion of the image based on a set of feature plane summations, and the feature vector is compared (810) with a corresponding object classifier to detect a presence of the corresponding object in the image portion of the image.