VGF-Filter - High-Speed Detection

High-Speed Spatial–Temporal Saliency Model for Infrared Small Moving Target Detection
A Aliha, Y Liu, G Zhou, Y Hu
Published in Remote Sensing, May 2024
DOI

Project Overview

Challenge: Real-Time Requirements

Practical deployment requires real-time processing:

• High frame rates (25-60 fps)
• Limited computational resources
• Low latency for time-critical applications
• Energy efficiency for embedded systems

Traditional methods are too slow; optimization needed.

Solution: Vectorized Guided Filter

VGF-Filter accelerates detection through:

1. Vectorized operations using SIMD instructions
2. Spatial-temporal saliency fusion
3. Optimized memory access patterns

Key Innovations

1. Vectorized Guided Filter

SIMD Optimization:

• Process 4-8 pixels simultaneously using AVX2
• Parallel sliding window computation
• Cache-friendly memory layout

Speed improvement: 10-50x faster than naive implementation

2. Spatial-Temporal Saliency

Spatial saliency: Local contrast in single frame Temporal saliency: Motion consistency across frames Fusion: Adaptive weighting based on scene dynamics

3. Multi-threaded Processing

Parallelization strategies:

• Frame-level parallelism
• Patch-level parallelism
• Optimized thread scheduling

Technical Details

Vectorization Example

// Vectorized box filter
__m256 vec_sum = _mm256_loadu_ps(&input[i]);
vec_sum = _mm256_add_ps(vec_sum, _mm256_loadu_ps(&input[i+1]));
// ... accumulate window
_mm256_storeu_ps(&output[i], vec_sum);

Processing Pipeline

Input Frame Sequence
    ↓
[Vectorized Preprocessing]
├── Parallel filtering
├── SIMD optimization
└── Memory-efficient access
    ↓
[Spatial-Temporal Saliency]
├── Spatial contrast computation
├── Temporal motion detection
└── Adaptive fusion
    ↓
[Fast Detection]
├── Thresholding
├── Connected component analysis
└── Target extraction
    ↓
Real-time Output

Results

Performance on CPU (Intel i7):

• Processing speed: 40+ fps
• Latency: < 25ms
• Accuracy maintained at 95%+ of non-optimized version

Platform compatibility:

• Desktop CPUs
• Embedded ARM processors
• GPU acceleration optional

Applications

Perfect for resource-constrained systems:

• Real-time surveillance cameras
• UAV onboard processing
• Battery-powered sensors
• Edge computing devices

Citation

@article{aliha2024high,
  title={High-Speed Spatial--Temporal Saliency Model: A Novel Detection Method for Infrared Small Moving Targets Based on a Vectorized Guided Filter},
  author={Aliha, A and Liu, Y and Zhou, G and Hu, Y},
  journal={Remote Sensing},
  volume={16},
  number={10},
  pages={1685},
  year={2024},
  publisher={MDPI}
}