SatTrack - Satellite Video Object Tracking

Single Object Tracking in Satellite Videos Based on Feature Enhancement and Multi-Level Matching Strategy
J Yang, Z Pan, Y Liu, B Niu, B Lei
Published in Remote Sensing, August 2023
DOI

Project Overview

Unique Challenges in Satellite Video Tracking

Satellite video tracking presents distinct challenges compared to conventional video:

• Large search space: Wide field of view requires searching large areas
• Dramatic scale variation: Target size changes drastically as it moves
• Limited texture: Small objects with few visual features
• Complex backgrounds: Urban environments with similar-looking objects
• Atmospheric effects: Clouds, haze, and lighting variations

Solution: Feature Enhancement + Multi-Level Matching

SatTrack combines:

1. Dual attention mechanism for discriminative feature enhancement
2. Hierarchical matching strategy for robust localization
3. Satellite-specific adaptations for optimal performance

Key Innovations

1. Channel-Spatial Attention Feature Enhancement

Problem: Traditional features lose discriminative power for small satellite targets

Solution: Dual attention module

• Channel attention: Automatically select informative feature channels
• Spatial attention: Focus on target regions, suppress background

Benefits:

• Enhanced target representation
• Suppressed background interference
• Improved discrimination capability

2. Multi-Level Matching Strategy

Three-level progressive refinement:

Adaptive transition between levels based on confidence scores

3. Scale-Adaptive Template Update

Challenge: Target appearance changes significantly with scale

Solution:

• Maintain template pyramid at multiple scales
• Online adaptation based on tracking confidence
• Historical template memory for occlusion recovery

Technical Architecture

Network Design

class SatTrack:
    def __init__(self):
        self.feature_enhancer = DualAttentionModule()
        self.multi_level_matcher = HierarchicalMatcher()
        self.scale_estimator = ScaleAdaptiveModule()
    
    def track(self, frame, template):
        # Feature enhancement
        enhanced_frame = self.feature_enhancer(frame)
        enhanced_template = self.feature_enhancer(template)
        
        # Multi-level matching
        for level in [1, 2, 3]:
            search_region = get_search_region(level, prev_location)
            response_map = self.multi_level_matcher(
                enhanced_frame[search_region], 
                enhanced_template
            )
            location = peak_location(response_map)
            
            if confidence_sufficient(response_map):
                break
        
        # Scale estimation
        scale = self.scale_estimator(frame, location)
        
        return location, scale

Experimental Results

Benchmark Dataset

SatSOT dataset evaluation:

• 100+ satellite video sequences
• Multiple target types: vehicles, vessels, aircraft
• Challenging scenarios: scale variation, occlusion, fast motion

Performance Highlights

Superior tracking accuracy:

• Outperforms generic trackers (SiamFC, SiamRPN) adapted to satellite videos
• Robust to scale changes up to 10x
• Effective in cluttered urban environments

Qualitative strengths:

• Handles large scale variations
• Recovers from temporary occlusions
• Distinguishes similar-looking objects
• Stable tracking over long sequences

Applications

• Traffic monitoring: Vehicle tracking on highways and urban roads
• Maritime surveillance: Vessel tracking in coastal waters
• Aircraft tracking: Airport and airspace monitoring
• Disaster response: Search and rescue operations
• Security applications: Critical area monitoring

Citation

@article{yang2023single,
  title={Single Object Tracking in Satellite Videos Based on Feature Enhancement and Multi-Level Matching Strategy},
  author={Yang, J and Pan, Z and Liu, Y and Niu, B and Lei, B},
  journal={Remote Sensing},
  volume={15},
  number={17},
  pages={4351},
  year={2023},
  publisher={MDPI}
}