VisionInfer

Lightweight Visual Language Model (VLM) Inference Tool optimized for Jetson Edge Devices and x86 platforms. Supports real-time inference for USB/RTSP cameras, VOD videos, and live streams with motion detection, frame deduplication, and efficient resource management.

Features

• 🎥 Multi-source support: USB cameras, RTSP streams, VOD files, live network streams
• 🚀 Motion-gated inference (only run inference when motion detected)
• 🎯 Frame deduplication (skip similar frames via L2 feature comparison)
• 📊 Real-time performance monitoring (encoding/inference time, frame metrics)
• 🔧 Jetson-optimized: Tailored for ARM64 architecture and limited edge resources
• 🎛️ Configurable parameters: Compression quality, inference interval, motion threshold
• 🪵 Debug mode for troubleshooting (--debug flag)

Requirements

General Requirements

• Python 3.8+
• OpenCV (cv2)
• NumPy
• psutil
• Ollama (v0.1.40+) [Optional]
• YOLO [Optional]
• FFmpeg (for frame extraction from streams/files)

Jetson-Specific Requirements

• Jetson Nano/Xavier NX/Orin (JetPack 6.0+)
• Minimum 8GB RAM

Installation

Install Dependencies Script Usage

Our install_deps.sh script supports flexible dependency installation with optional Ollama backend, and is compatible with both sh (dash) and bash on Ubuntu/Jetson systems.

Basic Usage

Scenario	Command
Install only core dependencies (ffmpeg, python3-pip, pipx)	curl -fsSL https://raw.githubusercontent.com/machinefi/VisionInfer/refs/heads/main/install_deps.sh | sh
Install core dependencies + Ollama backend	curl -fsSL https://raw.githubusercontent.com/machinefi/VisionInfer/refs/heads/main/install_deps.sh | sh -s -- --backend ollama
Show script help (check parameters)	curl -fsSL https://raw.githubusercontent.com/machinefi/VisionInfer/refs/heads/main/install_deps.sh | sh -s -- --help

Compatibility Note

• For better compatibility (especially on Jetson), you can replace sh with bash (recommended):

  # Install core dependencies + Ollama (bash execution)
  curl -fsSL https://raw.githubusercontent.com/iloveyou-github/VisionInfer/main/install_deps.sh | bash -s -- --backend ollama
  
  
  

Install VisionInfer

For Jetson (Pre-installed System OpenCV with CUDA)

To avoid breaking system dependencies (e.g., JetPack's pre-built OpenCV), use --system-site-packages to reuse the system's OpenCV:

• If you do not plan to use YOLO models in the future, we recommend installing using the following command.

pipx install --system-site-packages vinfer

• If you plan to use YOLO models in the future, we strongly recommend installing them with the following command.

pip install --system-site-packages vinfer
pip install ultralytics --no-deps
pip install matplotlib pillow polars psutil pyyaml requests scipy ultralytics-thop

For Other Systems (No Special OpenCV)

Install with full dependencies (includes OpenCV) if your system doesn't have a pre-configured OpenCV:

pipx install vinfer[full]

Jetson Resource Configuration

Increase Swap Space [Optional]

# Create 4GB swap file
sudo fallocate -l 4G /swapfile
sudo chmod 600 /swapfile
sudo mkswap /swapfile
sudo swapon /swapfile

# Make swap permanent (survive reboot) [Optional]
echo '/swapfile none swap sw 0 0' | sudo tee -a /etc/fstab

Configure GPU Memory (Jetson Orin/Nano)

# For Jetson Orin (set 16GB GPU memory)
sudo nvpmodel -m 0
sudo jetson_clocks

# For Jetson Nano (set max performance mode)
sudo nvpmodel -m 0
sudo jetson_clocks

Pull Optimized Model (Jetson)

# Recommended lightweight model for Jetson
ollama pull qwen3.5:2b

Quick Start

USB Camera Inference

# Basic USB camera (device ID 0) with debug logs
vinfer cam --usb-dev 0 --debug

# USB camera with motion detection (infer only on motion)
vinfer cam --usb-dev 0 --motion-gate --motion-threshold 500

# USB camera with frame deduplication (skip similar frames)
vinfer cam --usb-dev 0 --dedup --interval 2.0

# Basic USB camera (device ID 0) with YOLO
vinfer cam --usb-dev 0 --model "yolo"

# Basic USB camera (device ID 0) with YOLO26 in Detection task 
vinfer cam --usb-dev 0 --model "yolo" --yolo-version 26 --yolo-task "detection"

RTSP Camera Inference

# Basic RTSP stream (default credentials)
vinfer cam --rtsp-host 192.168.1.10 --rtsp-user admin --rtsp-pass password --debug

# RTSP with custom compression (320x240) and JPG quality (80)
vinfer --rtsp-host 192.168.1.10 --compress-size 320x240 --jpg-quality 80

# Simple RTSP stream (default credentials) with YOLO
vinfer -H 192.168.1.10 -m "yolo" 

# Simple RTSP stream (default credentials) with YOLO11 in Pose task
vinfer -H 192.168.1.10 -m "yolo" -yv 11 -yt "pose"

VOD (Video File) Analysis

# Local video file (analyze every 30 frames)
vinfer analyze --type vod --file /path/to/video.mp4 --start 0 --step 30

# Network VOD URL (e.g., MP4 stream)
vinfer analyze --type vod --url https://example.com/video.mp4 --debug

Live Stream Analysis

# HLS live stream (e.g., .m3u8)
vinfer analyze --type live --url https://example.com/stream.m3u8 --interval 1.0

Command Reference

Core Subcommands

Subcommand	Description
cam	Real-time camera inference (USB/RTSP)
analyze	Offline video/live stream analysis

Common Arguments

Argument	Short	Description	Default
--model	-m	Ollama model name or YOLO	qwen3.5:2b
--compress-size	-s	Frame compression resolution (WxH)	480x360
--jpg-quality	-q	JPG compression quality (0-100)	70
--motion-gate	-g	Enable motion detection (infer only on motion)	False
--motion-threshold	-T	Minimum motion area (pixels)	500
--dedup	-D	Enable frame deduplication (disabled if motion-gate is on)	False
--interval	-i	Inference interval (seconds/frame)	1.0
--debug	-d	Enable verbose debug logging	False
--Prompt	-r	User-defined prompts
--accelerate	-a	Accelerate reasoning speed	False
--version	-v	Show vinfer version
--yolo-version	-yv	Use YOLO version [8, 11, 26]	8
--yolo-task	-yt	Use YOLO task ['detection', 'segment', 'classify', 'pose', 'obb']	detection

Cam Subcommand Arguments

Argument	Short	Description
--rtsp-host	-H	RTSP server IP/domain (enables RTSP mode)
--rtsp-user	-U	RTSP authentication username
--rtsp-pass	-P	RTSP authentication password
--usb-dev	-u	USB camera device ID (0 = /dev/video0)
--show-preview	-p	Start live preview window

Analyze Subcommand Arguments

Argument	Short	Description
--type	-t	Analysis type (vod/live)
--file	-f	Local VOD file path
--url	-u	Network VOD/live stream URL
--start	-st	Start frame number (0-based)
--step	-sp	Inference frame interval

Troubleshooting

Common Issues & Solutions

Cannot uninstall sympy

• Symptom: Cannot uninstall Sympy 1.9

• Solution:

  sudo apt remove python3-sympy -y
  

numpy version conflict

• Symptom：numpy version conflict

• Solution:

  • Install the specified version

    sudo pip3 install numpy==1.23.5
    

EOF Error During Frame Extraction

• Symptom: EOFError/IOError when reading frames from RTSP/live streams
• Solutions:
  • Increase RTSP timeout: Add -stimeout 20000000 to FFmpeg command (code already includes this)
  • Check network stability (RTSP streams require low latency)
  • Use TCP for RTSP: --rtsp-transport tcp (enabled by default in code)

Zombie Processes (FFmpeg/Ollama)

• Symptom: Orphaned FFmpeg/Ollama processes consuming resources
• Solutions:
  • The code includes kill_all_ffmpeg() and stop_ollama_serve() for cleanup
  • Manually kill zombie processes:

    # Kill all FFmpeg processes
    sudo pkill -f ffmpeg
    
    # Restart Ollama service
    sudo systemctl restart ollama

Resource Exhaustion (Jetson)

• Symptom: Out of memory errors or slow inference
• Solutions:
  • Use smaller models (qwen3.5:2b instead of 7b)
  • Increase swap space (see Installation > Jetson Configuration)
  • Reduce frame resolution (--compress-size 320x240)
  • Increase inference interval (--interval 2.0 or higher)

Frame Extraction Failure

• Symptom: Frame extraction failed, unable to perform inference
• Solutions:
  • Verify RTSP URL/USB device accessibility
  • Check FFmpeg installation (ffmpeg -version)
  • For RTSP: Ensure camera is online and credentials are correct

Continuous Inference Errors

• Symptom: Continuous inference exception: [error message]
• Solutions:
  • Enable debug mode (--debug) to see detailed error logs
  • Check Ollama service status (sudo systemctl status ollama)
  • Verify model is pulled (ollama list to check installed models)

Known Limitations

Jetson-Specific Limitations

• Model Size: Avoid 7B+ models (e.g., qwen3.5:7b) on Jetson Nano/Xavier NX—use qwen3.5:2b for stable performance
• Inference Speed: 2B models run at ~1-2 FPS on Jetson Orin, ~0.5 FPS on Jetson Nano
• Preview Window: May be slow on Jetson Nano (disable with --no-preview if needed)

General Limitations

• RTSP Latency: RTSP streams may have 1-3s latency (normal for TCP transport)
• Frame Deduplication: May skip valid frames in low-motion scenarios (adjust DEDUP_THRESHOLD if needed)
• Motion Detection: Sensitive to lighting changes (tune --motion-threshold for your environment)

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• [YOLO](Ultralytics | Revolutionizing the World of Computer Vision) for end-to-end computer vision platform
• Ollama for lightweight LLM inference
• OpenCV for computer vision processing
• NVIDIA Jetson for edge AI platform support