Object Tracking via Bidirectional Label Propagation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object tracking technologies face difficulties in recognizing and tracking objects, especially in uncommon applications like fall recognition or bed motion recognition, as they require extensive manual data collection for accurate recognition.
Innovation Solution
A method and computer system that automatically tracks a target object by generating ground truth labels from first and last images in an image sequence, performing forward and backward tracking, and comparing results to generate a final tracking result, thereby speeding up the tracking process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual data collection and labeling is performed for object tracking in uncommon applications, then recognition accuracy is improved, but time consumption and labor cost increase significantly
Solution Approach 1:
The system performs preliminary actions by automatically generating ground truth labels for intermediate frames through forward and backward tracking from sparsely labeled key frames. This preliminary automated labeling reduces the need for extensive manual data collection while maintaining sufficient accuracy for uncommon applications like fall recognition.
Solution Approach 2:
The system creates copies of tracking information by propagating ground truth labels from labeled key frames to unlabeled intermediate frames through forward and backward tracking. This copying mechanism allows the system to generate sufficient training data without manual labeling of every frame, significantly reducing time consumption.
2Measurement precision
If extensive manual labeling is performed on all frames, then tracking accuracy is improved, but productivity decreases due to high labor cost
Solution Approach 1:
The system performs preliminary automated tracking by generating ground truth labels for intermediate frames through forward and backward tracking algorithms. This preliminary action produces sufficient tracking accuracy for training deep learning models without requiring manual labeling of every frame, thereby maintaining high productivity.
Solution Approach 2:
Instead of performing complete manual labeling on all frames (excessive action), the system applies partial automated tracking on key frames and propagates labels to intermediate frames. This partial action approach achieves sufficient tracking accuracy while dramatically improving productivity by reducing manual labor.
3Reliability
If forward and backward tracking are performed with multiple ground truth labels, then tracking reliability is improved, but computational complexity increases
Solution Approach 1:
The system segments the tracking process into two independent directional components: forward tracking from the first key frame and backward tracking from the last key frame. This segmentation allows the system to perform multiple ground truth label comparisons without requiring a single complex omnidirectional tracking algorithm, managing computational complexity through modular processing.
Solution Approach 2:
The system adds a temporal dimension to the tracking process by performing tracking in both forward (time-positive) and backward (time-negative) directions. This dimensional approach allows the system to generate multiple ground truth labels for verification, improving tracking reliability while maintaining manageable computational complexity through parallel processing.
Data Source
AI summary
A method and a computer system for object tracking are provided. The method includes the following steps. An image sequence including multiple images with a target object is obtained. A labelling operation corresponding to the target object in first two images and last two images in the image sequence is received to respectively generate four ground truth labels of the target object. A forward tracking of the target object is performed on the image sequence in time series according to the ground truth labels of the first two images to obtain a forward tracking result. A backward tracking of the target object is performed on the image sequence in time series according to the ground truth labels of the last two images to obtain a backward tracking result. The forward and the backward tracking results are compared to accordingly generate a final tracking result of the target object.