Features & Capabilities

🧠 Advanced AI Inference

At the heart of the system is the YOLOv11 (You Only Look Once) neural network, optimized via NVIDIA TensorRT.

• Classes Detected: Horse, Person.
• Precision: FP32 (High precision) or FP16 (High speed).
• Dynamic Batching: Automatically groups frames from multiple cameras to maximize GPU throughput.

📹 Intelligent Stream Management

The Camera Manager handles the complexity of RTSP streams:

• Auto-Reconnect: Automatically recovers lost streams without restarting the service.
• Hardware Decoding: Uses h264_cuvid to offload video decoding to the GPU media engine (NVDEC), sparing the CPU.
• Adaptive Framerate: Dynamically adjusts processing FPS based on system load.

📉 Motion Analysis & Behavior

We don't just detect objects; we analyze movement over time.

• Agitation Index: Measures how much the horse moves within a time window.
• Stressed State: Triggers alerts if agitation exceeds a threshold for >30 seconds.
• Resting State: Detects periods of inactivity to track sleep quality.

🌐 Distributed Design

The system is built on a Microservices Architecture:

• Decoupled: Camera ingestion is separate from AI inference.
• Resilient: If the AI crashes, the camera streams stay alive.
• Observable: Full Prometheus/Grafana integration for deep system visibility.