Target Position Tracking Without Manual Image Annotation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for obtaining the position of a target in multiple frames of images require manual annotation of real positions, leading to high labor costs and inefficient image processing.
Innovation Solution
A method utilizing machine learning and computer vision technologies to train a position obtaining model without manual annotation, using unsupervised learning to determine target positions in multiple frames of images, employing an initial model that adjusts its parameters based on error values to improve accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation of real target positions is performed in each frame of sample image, then the target tracking algorithm can be trained with accurate ground truth data, but the labor cost increases and the image processing process becomes cumbersome
Solution Approach 1:
The system performs self-service by using the target tracking algorithm itself to generate pseudo-labels for training data. The algorithm processes sample images to automatically determine target positions, creating training data without human intervention. This self-generated data then retrains and improves the algorithm, forming a closed-loop system that eliminates manual annotation while maintaining continuous improvement of tracking accuracy.
2Measurement precision
If manual annotation of real target positions is performed in each frame of sample image, then the target tracking algorithm can be trained with accurate ground truth data, but the labor cost increases
Solution Approach 1:
The system eliminates human labor by implementing self-service through automated pseudo-label generation. The target tracking algorithm automatically processes sample images to generate training data with target position information, completely replacing the need for manual annotation by human workers while maintaining the ability to produce accurate training data.
Solution Approach 2:
The system creates copies of target position information through pseudo-labels generated by the algorithm. Instead of requiring original manual annotations, the system produces replicated position data from processed images that can be used for training, effectively substituting human-created labels with algorithm-generated equivalents.
3Reliability
If traditional target tracking algorithm is used with manual annotation, then the target position can be obtained, but the overall process becomes complex and time-consuming
Solution Approach 1:
The system merges the target tracking algorithm with the data generation process into a single integrated workflow. The algorithm both determines target positions and generates training data simultaneously, combining what were previously separate steps (manual annotation and algorithm execution) into one automated process, thereby reducing overall process complexity.
Solution Approach 2:
The system implements continuous improvement by repeatedly processing images through the tracking algorithm and using the results to retrain and enhance the algorithm. This continuous cycle of processing and self-improvement maintains reliable target position determination while automating the entire process, eliminating the need for intermittent manual intervention and reducing overall process complexity.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
Disclosed are a method and apparatus for acquiring the positions of a target, and a computer device and a storage medium, belonging to the technical field of computers. The method comprises: acquiring a plurality of frames of images; calling a position acquisition model, wherein a model parameter of the position acquisition model is obtained by means of training on the basis of a first position of a selected target in a first sample image from among a plurality of frames of sample images, and a second position of the selected target in the first sample image; and by means of the position acquisition model and on the basis of the model parameter and the position of a target to be detected in a first image, determining the position of the target to be detected in a second image. According to the present invention, the plurality of frames of images are processed by means of the position acquisition model obtained by means of training, so as to acquire the positions of the target to be detected in the plurality of frames of images; and the position acquisition model can train an initial model by means of forward and reverse processes, without the need for related technical staff to perform manual annotation, thereby reducing labor costs, and improving the efficiency of the whole process of acquiring the positions of the target.