Pedestrian Tracking via Stereo Disparity and Body Part Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current object tracking technologies face challenges in accurately tracking pedestrians, especially when their posture changes or they are partially shielded, as existing methods struggle to maintain recognition in crowded scenes.

Innovation Solution

The method employs stereo cameras to generate disparity maps, extracts conspicuous portions of pedestrians using clustering algorithms, and calculates similarity between frames to track body parts, even when the whole body or head is not clearly visible, by utilizing disparity and grayscale histograms, and adjusting for deformation and positional relations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If whole-body templates are used for pedestrian detection, then detection can be performed when the pedestrian is fully visible, but detection accuracy deteriorates when the pedestrian is shielded or posture changes

Engineering Contradiction:
Improvepedestrian detection accuracyVSAvoidrobustness to shielding and posture changes
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent divides the pedestrian detection task into multiple body part detections (head, upper body, lower body) instead of detecting the whole pedestrian. Each body part has its own detection template, allowing the system to track pedestrians even when partially shielded by focusing on visible body parts rather than requiring the entire body to be visible.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimension of detection by using disparity maps in addition to grayscale images. This stereo vision approach provides depth information that helps distinguish pedestrians from background and handles occlusions better by detecting body parts at different depths.

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

2Measurement precision

If head-only detection is used when whole-body is not recognized, then detection can be performed in some occluded cases, but detection fails when head is not clear but other body parts are visible

Engineering Contradiction:
Improvepedestrian detection accuracyVSAvoiddetection coverage across different visibility conditions
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The system segments pedestrian detection into multiple body part detections (head, upper body, lower body). When the head is not visible, the system can still detect and track the pedestrian using upper body or lower body parts. Each body part serves as an independent detection target, providing redundancy that ensures detection continues even when specific body parts are occluded.

Inventive Principle:
Principle #1Segmentation

3Reliability

If traditional object tracking is used, then tracking works for static or simple moving objects, but tracking accuracy deteriorates when pedestrian posture frequently changes

Engineering Contradiction:
Improvetracking continuityVSAvoidtracking accuracy under posture changes
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

By segmenting the pedestrian into multiple body parts, the system can track each part independently. When posture changes occur, some body parts may change position or visibility, but others remain detectable. The system maintains tracking continuity by following the detectable body parts, ensuring reliable tracking even during posture transitions.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system dynamically adapts to posture changes by flexibly selecting which body parts to track based on current visibility. The tracking algorithm adjusts to the dynamic nature of pedestrian movement, maintaining accurate tracking by following the most visible or stable body parts at each moment rather than requiring a fixed whole-body template.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3032462B1Method and apparatus for tracking object, and non-transitory computer-readable recording medium
Publication Date: 2020.02.26 RICOH CO LTD
  • EP3032462B1 patent drawingFigure 1
  • EP3032462B1 patent drawingFigure 2
  • EP3032462B1 patent drawingFigure 3A~3B

AI summary

A method and an apparatus for tracking an object are disclosed. The method includes extracting conspicuous portions of the object of a current frame; performing clustering for disparity points of a next frame to form clusters; calculating similarity between each cluster and the conspicuous portion to determine candidate clusters similar to the conspicuous portion; classifying the candidate clusters based on positional relation among the conspicuous portions; determining similarity between the group and the object based on the similarity between the candidate clusters and the conspicuous portion; and selecting the group with the highest similarity and taking a position of the selected group as a position of the object. According to the object tracking method, an accurate tracking result can be obtained, even when a posture of the object frequently changes or the object is partially shielded.