Image Object Tracking With Adaptive Similarity Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object tracking techniques using deep neural networks face challenges with tracking accuracy due to similar objects and variations in object appearance, leading to tracking errors.
Innovation Solution
An information processing apparatus that detects a collation target from images, calculates an evaluation value based on feature values, and adjusts a similarity threshold using distribution statistics to enhance tracking accuracy by reducing matching errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep neural network is used for object tracking, then tracking accuracy is improved, but tracking errors occur when similar objects or objects with varying appearances are detected
Solution Approach 1:
The patent applies dynamics by making the similarity threshold adaptive rather than fixed. The threshold is dynamically adjusted based on the distribution of evaluation values obtained from comparing objects across frames. When objects have varying appearances or similar objects are detected, the system automatically adjusts the threshold to maintain reliable tracking, resolving the contradiction between precision and reliability.
Solution Approach 2:
The patent changes the parameter of the similarity threshold based on statistical distribution of evaluation values. By calculating the threshold according to the distribution characteristics (mean, standard deviation) of evaluation values from previous frames, the system adapts to different tracking conditions and maintains both high accuracy and reliability across varying scenarios.
2Productivity
If a fixed similarity threshold is used for object matching, then processing speed is maintained, but tracking accuracy deteriorates when object appearance varies
Solution Approach 1:
The system transitions from a static fixed threshold to a dynamic adaptive threshold that changes based on the distribution of evaluation values in real-time. This allows the threshold to automatically adjust to matching conditions while maintaining efficient processing, as the adjustment is based on statistical calculations from previous frame data.
Solution Approach 2:
The patent performs preliminary calculation of evaluation value distributions from previous frames to determine the threshold for current frame processing. By preparing the threshold in advance based on historical data, the system ensures accurate tracking without adding significant computational delay to the main processing pipeline.
Data Source
AI summary
An information processing apparatus detects a collation target from an image and obtains a predetermined evaluation value between the collation target previously detected and the collation target detected from the image. The information processing apparatus performs collation processing on the collation target detected from the image and the collation target previously detected based on the evaluation value obtained for the collation target detected from the image and a distribution of evaluation values previously obtained.