Correlation Filter Target Tracking via Average Pooling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing correlation filter-based target tracking algorithms face challenges with parameter over-fitting due to limited training data, leading to imbalanced quantities of parameters and samples, which affects tracking precision.
Innovation Solution
A target tracking method that employs an average pooling operation on filter weights corresponding to different feature channels of an input feature map, reducing the number of algorithm parameters while maintaining the quantity of training sample features, thereby avoiding parameter over-fitting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If correlation filter-based target tracking algorithm is used, then tracking operation can be performed, but parameter over-fitting occurs due to large number of parameters and limited training samples
Solution Approach 1:
The patent merges multiple feature channels into a unified correlation filter model, where filter weights are shared across different feature channels. This combining approach reduces the total number of parameters while maintaining tracking functionality across multiple feature representations, directly addressing the parameter over-fitting problem.
Solution Approach 2:
The correlation filter is designed to be universal across multiple feature channels, with a single set of filter weights applied to all channels. This multi-functional design allows the same filter to operate on different feature representations simultaneously, reducing parameter redundancy and preventing over-fitting while maintaining tracking precision.
2Device complexity
If ordinary pooling operation is used to reduce parameters, then parameter quantity decreases, but quantity of samples is also reduced which cannot resolve parameter-sample imbalance
Solution Approach 1:
The patent segments the feature processing into two distinct stages: feature extraction (which preserves all sample information) and filter weight sharing (which reduces parameters). By separating these functions, the method reduces parameter quantity without compromising the quantity or quality of training samples, unlike ordinary pooling that loses sample information.
Solution Approach 2:
The correlation filter weights serve as an intermediary that bridges multiple feature channels. Instead of reducing sample quantity through pooling, the shared weights act as a mediator that enables parameter reduction while maintaining full sample information across all feature channels, resolving the parameter-sample imbalance.
Data Source
Figure 1
Figure 2~3
Figure 4~5
AI summary
The present application provides a target tracking method, comprising: obtaining a plurality of consecutive picture frames of a target video, setting a tracking target area of the nth picture frame in the plurality of consecutive picture frames; performing a feature extraction operation on the tracking target area of the nth picture frame, to obtain an input feature map of the tracking target area; calculating a weighting filter corresponding to the input feature map of the tracking target area according to a correlation filtering algorithm and an average pooling constraint condition; calculating, by using the weighting filter and the input feature map of the (n+1)th picture frame in the plurality of consecutive picture frames, an output response of an (n+1)th picture frame; and confirming a tracking target area of the (n+1)th picture frame according to the output response of the (n+1)th picture frame, and performing operations until the tracking target areas of all the consecutive picture frames are obtained. The present application further provides a target tracking apparatus. The present application improves the precision and the effectiveness of target tracking.