Shared-Backbone Object Tracking for Detection and Re-Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection and tracking methods in autonomous driving technologies suffer from performance degradation when using separate networks for object detection and re-identification, leading to slower tracking algorithms.
Innovation Solution
A method and apparatus that utilize a knowledge distillation technique to train a single network with a shared backbone for both object detection and re-identification, enabling simultaneous performance of these tasks through multi-task learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If separate networks are used for object detection and re-identification, then each network can be optimized for its specific task, but the overall tracking algorithm speed decreases and system complexity increases
Solution Approach 1:
The patent merges object detection and re-identification into a single unified network architecture. The detection model includes a backbone network that extracts features from input images, a first neck that processes detection information, and a second neck that generates feature vectors for re-identification. This integration eliminates the need for separate networks while maintaining both detection and re-identification capabilities, thereby improving tracking algorithm speed.
Solution Approach 2:
The unified detection model performs multiple functions through its different components: the backbone network handles feature extraction for both detection and re-identification, the first neck processes detection information, and the second neck generates re-identification feature vectors. This multi-functionality allows the system to achieve both task-specific optimization and improved processing speed.
2Productivity
If a single unified network is used for both object detection and re-identification, then tracking algorithm speed improves, but the network complexity and training difficulty increase
Solution Approach 1:
The unified network is segmented into distinct functional modules: a backbone network for feature extraction, a first neck for detection information processing, and a second neck for re-identification feature vector generation. This segmentation allows each module to be optimized for its specific function while maintaining overall system integration, managing complexity through modular design.
3Productivity
If knowledge distillation is applied to train the re-identification network, then training efficiency improves, but the complexity of the training process increases
Solution Approach 1:
The patent employs knowledge distillation where a pre-trained detection model serves as a teacher model to guide the training of the re-identification network. The teacher model's output (detection information and feature vectors) is used as training data for the student re-identification network, enabling efficient transfer of knowledge and improving training efficiency through a structured intermediary training approach.
Data Source
AI summary
A tracking method according to an example of the present disclosure may include generating, by a generation device, a first feature based on a first frame through a backbone, generating, by the generation device, first detection information indicating a detection result for a first object based on the first feature through a first neck for object detection, generating, by the generation device, a first feature vector for a visual feature of the first object based on the first feature through a second neck for object re-identification, and/or performing, by a tracking device, tracking based on the first detection information and the first feature vector.