Video Object Identification via Motion Vector Consistency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning models for object identification in video streams face challenges in achieving real-time analytics without sacrificing accuracy, and the creation of large, accurate training datasets is hindered by the time-consuming process of manual labelling and the errors inherent in automatic labelling methods.
Innovation Solution
A computer-implemented method that identifies objects within a video stream by deriving a cumulative motion vector between temporally spaced frames to determine consistency, allowing for the generation of a dataset for training a machine learning-based object classifier, which can include optional features like overlapping camera views and threshold-based frame storage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labelling is used to create training datasets, then accuracy of object identification is improved, but time consumption and productivity are worsened
Solution Approach 1:
The patent uses automatically generated labels as an intermediary step in the dataset creation process. Rather than directly comparing manual labels with automatic labels, the system uses automatic labels to generate training datasets that are then refined through motion vector verification, creating a multi-stage intermediary process that balances speed and accuracy
Solution Approach 2:
The patent replaces the mechanical process of manual verification with an automated system using motion vectors and consistency checks. Instead of human operators manually verifying each label, the system uses computational methods to verify label consistency across frames, dramatically improving productivity while maintaining accuracy
2Productivity
If automatic labelling is used to create training datasets, then productivity is improved, but accuracy and reliability are worsened due to errors in automatic labelling
Solution Approach 1:
The patent implements feedback mechanisms where motion vectors are used to verify the consistency of automatic labels. The system checks whether objects identified in consecutive frames maintain consistent positions and identities according to their motion vectors, providing feedback that can identify and correct labelling errors
Solution Approach 2:
The patent performs preliminary verification of automatic labels using motion vector analysis before the labels are used for training. By checking consistency in advance, the system prevents erroneous labels from contaminating the training dataset, ensuring higher reliability
3Measurement precision
If large training datasets are created to improve model accuracy, then measurement precision is improved, but loss of time and computational resources are worsened
Solution Approach 1:
The patent performs preliminary filtering and verification of training examples using motion vector consistency checks before they are added to the training dataset. This preliminary action ensures that only high-quality, consistent examples are included, reducing the need for extensive training on large datasets and decreasing training time
Data Source
AI summary
A computer-implemented method of identifying an object within a video stream from a camera, and determining the consistency with which the object is identified within plural temporally spaced video frames of the video stream.