Multi-Target Tracking Feature Decoupling for Detection and Re-ID
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing joint detection and re-identification multi-target tracking methods face conflicts in optimization directions during the training process, leading to low tracking accuracy due to the inherent differences between object detection and re-identification tasks.
Innovation Solution
A multi-target tracking method with joint detection and re-identification based on feature decoupling, utilizing a channel and spatial decoupling approach to separate features for detection and re-identification tasks, employing a DLA-34 backbone network, channel decoupling units, and spatial decoupling units within a multi-target tracking model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If joint detection and re-identification multi-target tracking methods are used to reduce inference time, then productivity is improved, but tracking accuracy deteriorates due to conflicts in optimization directions between detection and re-identification tasks
Solution Approach 1:
The patent applies segmentation by dividing the feature extraction process into separate detection and re-identification branches. Each branch has its own feature extraction network optimized for its specific task, allowing independent optimization without conflict. The detection branch focuses on location and category while the re-identification branch focuses on appearance features, resolving the optimization direction conflict while maintaining joint processing efficiency.
2Device complexity
If a single backbone network is used for both detection and re-identification tasks, then device complexity is reduced, but tracking accuracy deteriorates due to conflicting optimization directions
Solution Approach 1:
The patent segments the single backbone network into multiple specialized branches. The detection branch extracts features optimized for object localization and classification, while the re-identification branch extracts features optimized for appearance-based identification. This segmentation allows each branch to be trained with task-specific loss functions, eliminating optimization conflicts while maintaining overall system efficiency.
Solution Approach 2:
The patent applies local quality by giving different parts of the network different functions and optimization targets. The detection branch uses loss functions focused on bounding box accuracy and class classification, while the re-identification branch uses loss functions focused on appearance feature discrimination. This localized optimization strategy improves tracking accuracy without significantly increasing overall device complexity.
3Measurement precision
If detection and tracking stages are separated in detection-based tracking, then tracking accuracy is improved through independent optimization, but device complexity increases due to requiring two computationally intensive components
Solution Approach 1:
The patent merges the detection and tracking stages into a unified joint detection and re-identification framework. By integrating these stages, the system achieves accuracy comparable to separate-stage methods while reducing overall system complexity. The unified architecture shares computational resources and eliminates the need for two independently optimized components, resolving the contradiction between accuracy and complexity.
Data Source
AI summary
A multi-target tracking method with joint detection and re-identification based on feature decoupling in the field of multi-target tracking for solving low tracking accuracy problem caused by conflicts in optimization directions in the joint training process of detection and re-identification tasks and improving tracking performance, includes the steps of: preprocessing multi-target tracking dataset, extracting feature F from the processed images and labels by the backbone network; decoupling feature F in the channel and spatial dimensions to obtain feature Fdet suitable for detection tasks and feature Fid suitable for re-identification tasks respectively, obtaining the position of the target detection bounding box in the image based on Fdet subdivision detection branch; calculating the similarity matrix D based on Fid and IOU overlap level, and processing matching of the target detection bounding box in the image with the trajectory already existed by using D and Hungarian algorithm to finally obtain the tracking result.