Mariclip: Real-Time Maritime Video Segment Extraction Via Edge-Optimized Detection and Tracking

Maritime surveillance systems, tasked with monitoring ships, buoys, and critical aquatic activities, generate vast amounts of video data that challenge manual analysis. While automated video clipping techniques can alleviate this burden, existing methods suffer from computational inefficiency or fail to preserve temporal continuity of target-specific events. To address these limitations, we propose MariClip, a lightweight trackingdriven framework optimized for real-time maritime video segment extraction on edge devices. At its core, MariClip integrates a YOLOv7-tiny detector, fine-tuned with maritime-specific augmentations (wave distortions, glare simulations), and a simplified Kalman filter tracker that operates on a 4D state vector (position + velocity). By executing full-frame detection sparsely (every 5th frame) and relying on motion prediction otherwise, our system reduces computational load by 79 % compared to per-frame detection approaches while maintaining temporal coherence. Adaptive clipping rules dynamically validate segments: initiation requires 3+ consecutive tracking confirmations to suppress false positives, while termination is triggered after <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$5+$</tex> tracking failures, with 1 -second buffers added to capture critical pre-/post-event contexts. Our framework MariClip achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 4. 7 \% ~ m A P @ 0. 5 ~ o n ~}$</tex> a static maritime dataset (13,200 images), surpassing YOLOv7tiny (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 8. 2 \%}$</tex>) and YOLOv5s (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 2. 1 \%}$</tex>). Qualitative video tests show stable tracking under waves/occlusions, with geometrybased categorization. Deployed on Jetson TX2, it runs at 26 FPS with much lower CPU load than per-frame detection.