Jetson Vision

Real time video analytics with Nvidia's Jetson devices.

Prerequisites

Hardware

• Tested with NVIDIA Jetson Nano (Jetpack 4.6 [L4T 32.6.1])

Docker Nvidia Runtime

sudo vim /etc/docker/daemon.json

{
    "runtimes": {
        "nvidia": {
            "path": "nvidia-container-runtime",
            "runtimeArgs": []
        }
    },
    "default-runtime": "nvidia"
}

sudo service docker restart

# Check
sudo docker info | grep Default

# Expected output
Default Runtime: nvidia
WARNING: No blkio weight support
WARNING: No blkio weight_device support

How to Start

Docker Enviornment

./scripts/build.sh

# Start docker container
./scripts/start.sh

Download Pre-trained Models

• Pretrained models will be downloaded @ /jetson-inference/data/networks
• In scripts/start.sh, the models directory is mounted to local volume (/media/data/models/). Thus, no need to re-download the models multiple times in docker environment.

cd /jetson-inference/tools
./download-models.sh

Basics

Hello world codes for using Jetson-inference.

cd basics

Object Detection

python3 detect.py /dev/vidoe0

Semantic Segmentation

python3 segment.py /dev/video0

Analytics

Using Jetson-inference toghether with Supervision to do vidoe analytics.

cd analytics/

People Counting in a Zone

Counting number of people in a defined polygon zone.

python3 counting.py /dev/video0

Flow analysis

Counting objects going in and going out of a line zone.

python3 flow.py /dev/video0

People Redaction

Detects person and redact the whole body, could be used to process video with privacy concerns.

python3 redaction.py /dev/video0