A Combined Frame Difference and Convolution Method for Moving Vehicle Detection in Satellite Videos

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Existing vehicle detection methods struggle with missed detections due to poor shape/texture features and blurred boundaries.
• Satellite video analysis presents unique challenges for object detection.

Purpose of the Study:

• To develop a novel moving vehicle detection approach for satellite videos.
• To improve detection accuracy by effectively integrating spatiotemporal information.

Main Methods:

• A frame difference module (FDM) combines frame difference and convolution to extract and refine motion features.
• A backbone network and neck with deformable convolution enhance spatiotemporal feature extraction and multiscale integration.
• A small-target self-reconstruction module (SR-TOD) improves small vehicle feature representation.

Main Results:

• The proposed method significantly reduces missed detections caused by weak features and blurred boundaries.
• Achieved notable improvements in satellite video moving vehicle detection.
• Demonstrated an average F1-score increase of 3.9% and a speed enhancement of 7s per frame over DSFNet.

Conclusions:

• The novel approach effectively addresses limitations in existing satellite video detection methods.
• The integration of spatiotemporal features and specialized modules enhances moving vehicle detection performance.