Template Matching for Eye Localization

Template matching image recognition serves as a fundamental processing method in the image processing domain, applicable to industrial scenarios including eye localization, character matching, and machine part identification. This method achieves spatial alignment of images captured by different sensors or the same sensor under varying conditions, or identifies corresponding patterns in target images based on known templates. Essentially, a template is a known small image, while template matching involves scanning a larger image to locate targets that share identical dimensions, orientation, and visual characteristics with the template. Through specific algorithms, targets can be identified with their coordinate positions determined.

Taking 8-bit images (where each pixel is represented by one byte) as an example, the template T (m × n pixels) is overlaid onto the search image S (W × H pixels) through sliding translation. The region covered by the template is called sub-image Sij, where i and j represent the coordinates of the sub-image's top-left corner in S. The search range is defined by:

1 ≤ i ≤ W – M

1 ≤ j ≤ H – N

The matching process is completed by comparing the similarity between template T and sub-image Sij. To quantify matching quality, two common metrics are used (code implementation typically employs either Sum of Absolute Differences or Normalized Cross-Correlation):

Expanding the first equation reveals three components: the third term represents the template's total energy (constant), the first term denotes sub-image energy under the template (slowly varying), and the second term is the cross-correlation function between sub-image and template (rapidly changing). When template T and sub-image Sij match perfectly, this term reaches its maximum value. Thus, the following correlation function can reflect matching degree:

2 Mathematical Modeling

2.1 Calculation Formulas

The template matching algorithm computes similarity between the template and matching regions, selecting the position with highest similarity as the match point. Given industrial computers' limited processing capacity, mathematical models must be simplified based on infrared thermal imaging tracking characteristics, prioritizing minimal computation. Target tracking algorithms determine target positions using relative matching error magnitudes as discrimination criteria—the position with minimum error indicates the target, regardless of absolute similarity. Therefore, Formula (1) with minimal computation and highest efficiency is selected as the original mathematical model. The matching point location algorithm searches for the point with minimum matching error across the entire region, expressed as Formula (6):

Where variables K and L represent matching region dimensions, while M and N denote template dimensions.

2.2 Template Size

Template size significantly impacts system performance and computational load. Oversized templates degrade dynamic characteristics, while undersized templates reduce target feature data, decreasing matching sensitivity and increasing detection difficulty. Practical experiments show optimal results with a template size of 32×16 pixels.

2.3 Matching Region

Different applications impose varying requirements on matching regions and real-time performance. Photoelectric detection devices must complete real-time data processing within video acquisition cycles (20ms). Since target movement between video frames is minimal with negligible feature changes, matching regions can be substantially reduced. Excessively small regions impair dynamic characteristics, while oversized regions drastically increase computation. Region size selection must balance device parameters. Since CCIR-standard video signals use interlaced scanning and systems process data field-by-field for real-time performance, resulting images maintain a 2:1 aspect ratio. To maintain consistent horizontal/vertical dynamic characteristics, matching regions follow the same 2:1 proportion. Selecting relatively large matching ranges while satisfying real-time requirements can enhance device performance.