Multi-model estimation based moving object detection for aerial video.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Unmanned Aerial Vehicle (UAV) technology is rapidly advancing, increasing the need for robust moving target detection in aerial videos.
• Existing methods often fail in complex backgrounds (e.g., viaducts, buildings, trees) due to single background model constraints.

Purpose of the Study:

• To develop an accurate moving target detection method for aerial videos in complex, multi-background environments.
• To overcome the limitations of existing registration-detection frameworks that struggle with intricate scenes.

Main Methods:

• Automatic estimation of multiple background models by segmenting scenes into color blocks and estimating dense optical flow.
• Calculating affine transformation models for large blocks and merging consistent models.
• Pixel-wise calculation of subordinate degrees to multi-background models for small blocks.
• Moving object segmentation using an energy optimization method solved via Graph Cuts.

Main Results:

• The proposed method effectively removes false alarms caused by static structures like buildings and trees.
• Accurate detection of moving objects in complex aerial video scenes was achieved.
• Experimental results on public aerial videos validate the effectiveness of the multi-background model approach.

Conclusions:

• The developed method successfully addresses the challenge of moving target detection in complex backgrounds.
• By analyzing pixel-wise relationships to multiple background models via energy minimization, the system achieves high accuracy.
• This approach significantly improves upon existing methods for aerial surveillance and analysis.