Multi-Part Target Tracking for Occlusion and Orientation Changes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing target tracking methods struggle with low recognition success rates and instability due to difficulty in recognizing targets across video frames, particularly when the target changes orientation or is partially obscured, leading to increased identifier IDs and loss of tracking.
Innovation Solution
A target tracking method that performs dual part recognition (first and second part) and uses a multi-target tracking model to enhance recognition and stability by enlarging the region of unmatched boxes, employing trained Faster RCNN models and bipartite graph matching for accurate box associations, and integrating Kalman filter trackers for improved tracking.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If face recognition is used to track targets in video frames, then target identification can be achieved when the target is facing forward, but tracking fails when the target turns its head or changes orientation
Solution Approach 1:
The patent divides the target recognition task into multiple parts: face recognition results are combined with body part recognition results (such as head, shoulders, torso). By segmenting the recognition into multiple components, the system can track targets even when the face is not visible, as long as other body parts are detectable.
Solution Approach 2:
The tracking system uses multiple recognition functions simultaneously - face recognition, body part recognition, and tracking by association. This multi-functional approach allows the system to adapt to different target orientations and occlusion conditions, maintaining tracking reliability across various scenarios.
2Measurement precision
If face recognition matching is used between sequential frames, then targets can be tracked when consistently visible, but the number of identifier IDs increases abnormally when recognition fails
Solution Approach 1:
The system uses feedback mechanisms to verify matching results. When face recognition fails to match a target across frames, the system checks body part recognition results and tracking history to determine whether to create a new identifier or associate with an existing one, preventing abnormal ID proliferation.
Solution Approach 2:
The system performs preliminary body part recognition and association before finalizing target identification. By pre-establishing relationships between body parts and potential targets, the system can more accurately determine whether a new identifier is needed, reducing false positives in ID creation.
3Ease of operation
If only face recognition boxes are used for tracking, then the tracking process is simple, but targets are lost when the face is not visible or recognizable
Solution Approach 1:
The patent extends tracking from a single dimension (face recognition) to multiple dimensions by incorporating body part recognition results. The system considers not only face boxes but also head, shoulders, and torso recognition results, creating a multi-dimensional tracking approach that maintains reliability when face recognition alone fails.
Data Source
AI summary
A target tracking method, including: performing first and second part recognitions in a current frame, to obtain recognition boxes of the first and second parts, respectively; performing a matching between the recognition boxes of the first and second parts, to obtain a first matching result, and determining a box to be tracked of the second part of a recognized target in the current frame, according to the first matching result; and performing a matching between a first multi-target tracking model of a previous frame and a respective box to be tracked of the second part in the current frame, to obtain a second matching result, and determining a target tracking result of the current frame according to the second matching result, where the first multi-target tracking model of the previous frame includes main trackers of a plurality of targets to be tracked in the previous frame.