|
Moving object extraction based on accumulative difference images and unit gradient vectors for gait image analysis |
|---|---|
| รหัสดีโอไอ | |
| Title | Moving object extraction based on accumulative difference images and unit gradient vectors for gait image analysis |
| Creator | Thanyamon Pattanapisont |
| Contributor | Waree Kongprawechnon, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | Moving object extraction, Background subtraction, Integrated approach for moving object extraction, Accumulative difference images, Unit gradient vectors |
| Abstract | Moving object extraction is a fundamental process in computer vision applications, particularly in appearance-based gait analysis, where the quality of the extracted walking-person region directly affects subsequent gait feature representation. The conventional Accumulative Difference Image (ADI) technique can effectively extract moving regions corresponding to walking individuals, however, its reliance on pixel intensity makes it sensitive to environmental variations. Intensity-based ADI (INT) generally provides high foreground coverage, but it is prone to False Positive (Type I) errors caused by shadows, floor reflections, illumination changes, and moving background regions. In contrast, gradient-based approaches, such as Unit Gradient Vector-based ADI (UGV), can suppress some illumination-related artifacts, but they are vulnerable to False Negative (Type II) errors when gradient or textural information is weak. This limitation often leads to fragmented or incomplete gait images, especially when the subject wears plain clothing or appears under low-contrast conditions.To address these limitations, this study proposes an improved walking-person extraction methodology that integrates intensity-based and gradient-based ADI. Two integration approaches are developed. The first approach is a pixel-wise switching method, which selects either INT or UGV according to local textural information measured by Standard Deviation (SD) or Gradient Value (GV). This approach provides an interpretable implement for analyzing condition-specific extraction behavior. The second approach is a pixel-wise merging method, which combines INT and UGV through a weighted arithmetic mean. Instead of switching between methods, this approach adjusts the contribution of both methods according to the availability of reliable gradient information, thereby improving foreground completeness while reducing environmental noise.Experimental evaluations were conducted using public gait datasets and author-collected datasets under various conditions, including normal walking, complex illumination, shadow, reflection, low light, low contrast, dynamic background, plain clothing, and view-angle variation. The results show that Approach 1 is useful for understanding the behavior of SD and GV under different scene conditions, but its switching decision can be sensitive to unreliable texture estimation. In contrast, Approach 2 provides the best overall performance and is concluded as the recommended method of this study. Quantitatively, Approach 2 improves the IoU score by 14.62% over INT and 65.13% over UGV. It also improves the F1-score by 9.41% and 42.70% over INT and UGV, respectively, and improves the F2-score by 3.52% and 60.94%. These findings indicate that the proposed merging strategy effectively reduces both False Positive and False Negative errors by combining the foreground coverage strength of INT with the noise-suppression capability of UGV.The proposed method also demonstrates a fail-safe capability when gradient information is insufficient. The system adaptively relies more on INT to prevent extraction failure. This behavior is particularly important in low-texture and plain-clothing scenarios where conventional UGV may fail. Overall, the proposed methodology improves the correctness and completeness of walking-person extraction and provides a more reliable foundation for subsequent gait image analysis in practical environments. |
