Moving Object Detection under a Moving Camera via Background Orientation Reconstruction.

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Moving object detection is crucial for applications like autonomous driving and surveillance.
• Complex backgrounds and camera motion significantly challenge existing detection algorithms.
• Accurate background motion estimation is essential for reliable foreground object identification.

Purpose of the Study:

• To propose a robust method for moving object detection under a moving camera.
• To reconstruct the background orientation field to differentiate static from dynamic elements.
• To enhance the saliency of moving objects and improve detection consistency.

Main Methods:

• Reconstruction of the background orientation field using Poisson fusion based on modified gradients.
• Calculation of a motion saliency map by comparing original and reconstructed orientation fields.
• Weighted accumulation enhancement considering appearance and motion similarity.
• Incorporation of motion continuity to reduce false positives.

Main Results:

• The proposed method effectively highlights motion saliency of moving objects.
• Improved consistency between object and background regions was achieved.
• Experimental results on public datasets demonstrate superior performance compared to state-of-the-art methods.
• The technique shows robustness in complex background scenarios.

Conclusions:

• The background orientation field reconstruction method offers a promising approach for moving object detection.
• Poisson fusion and weighted accumulation enhance detection accuracy and robustness.
• The method successfully addresses challenges posed by camera motion and complex backgrounds.