SkySentry ADS: Aerial Deterrent System

SkySentry ADS is a computer vision pipeline engineered for the real-time detection, classification, and tracking of small-profile aerial targets. The system is optimized for high-clutter environments where traditional video inference models may struggle.

Technical Overview

The pipeline utilizes a YOLOv11-based architecture fine-tuned on a dataset of avian profiles to minimize false-positive triggers.

Detection Logic & Optimization

To maintain high precision in long-range detection, the system implements Slicing Aided Hyper Inference (SAHI). This allows for the detection of "small" objects by partitioning the input frame $I$ into $n$ overlapping patches $P$.

The detection accuracy is evaluated via the Intersection over Union (IoU) metric,

$$IoU = \frac{Area(A \cap B)}{Area(A \cup B)}$$

where $A$ is the predicted bounding box and $B$ is the ground truth. SkySentry is tuned for a high $mAP_{50}$ (Mean Average Precision) to ensure reliability in various environments.

Metric	Target Specification	Implementation Logic
Detection Horizon	Small-profile targets ($<20 \text{px}$)	Slicing Aided Hyper Inference (SAHI): Divides 1080p/4K streams into overlapping $640 \times 640$ tensor batches to recover spatial resolution.
Inference Latency	$<15 \text{ms}$ (Backbone)	Quantized YOLOv11n: Optimized for edge deployment (MPS/CUDA).
Compute Efficiency	Dynamic Load Balancing	Motion-Vector Culling: Pre-processing step (cv2.absdiff) discards static sky/terrain tiles before GPU inference, reducing load by ~60% in stable environments.
Precision ($mAP_{50}$)	High-Recall Optimization	Non-Maximum Suppression (NMS): Custom intersection logic merges overlapping tile detections to prevent double-counting.

System Features

• Motion-Adaptive Inference: Implements a pre-processing motion vector check (cv2.absdiff) to cull static tiles (e.g., blue sky) prior to inference, significantly reducing GPU load on high-resolution streams.

• Automatic Range Profiling: run-time configuration manager (RangeProfileManager) that adjusts confidence thresholds ($0.05 - 0.95$) and SAHI tiling strategies based on target distance and environmental clutter.

• Hybrid Inference Modes: Supports automated switching between standard YOLOv11 full-frame inference (for close-range/large targets) and Batched SAHI (for long-range/small-profile targets) to optimize the latency-accuracy trade-off.

• Edge-Compute Optimization: Pipeline architecture designed for low-SWaP (Size, Weight, and Power) deployment, utilizing batch processing to maximize throughput on MPS (Apple Silicon) and CUDA backends.

Access Note

The full source code for SkySentry ADS is currently kept in a private repository for security and research integrity.