tonible14012002/fall_detector

Human Fall Detection On Jetson Xavier NX - Jetpack 5.1

Prerequisite

Jetpack 5.1 (prefer if run with cuda)
Python 3.8

Installation

1. Clone the repo

git clone git@github.com:tonible14012002/fall_detector.git

2. Install Dependencies

For cpu devices

pip install -r requirements.txt

For jetson devices

Install other dependencies

pip install -r requirements.txt

Uninstall pytorch and torchvision because previous command will install cpu only version of these packages.

pip uninstall torch torchvision

Installation for cuda enable devices required pytorch, torchvision to be compiled using cuda enabled. Follow this instruction to install compatible pytorch version .

If you are using jetpack, there might be preinstalled pytorch version.

Run

See Makefile for quick commands.

Quick start

python main.py -C '<path_to_your_video>'

Start with cpu device

python main.py -C '<path_to_your_video>' --device cpu

Jetson Xavier NX setup

Maximize Power Model Jetpack Version

> show nvidia-jetpack
# ---- Results -----
Package: nvidia-jetpack
Version: 5.0.2-b231
Priority: standard
Section: metapackages
Maintainer: NVIDIA Corporation
Installed-Size: 199 kB
Depends: nvidia-jetpack-runtime (= 5.0.2-b231), nvidia-jetpack-dev (= 5.0.2-b231)
Homepage: http://developer.nvidia.com/jetson
Download-Size: 29,3 kB
APT-Sources: https://repo.download.nvidia.com/jetson/common r35.1/main arm64 Packages
Description: NVIDIA Jetpack Meta Package

Checking supported power models

> nvpmodel -p --verbose | grep POWER_MODEL
# ---- Result ----
NVPM VERB: POWER_MODEL: ID=0 NAME=MODE_15W_2CORE
NVPM VERB: POWER_MODEL: ID=1 NAME=MODE_15W_4CORE
NVPM VERB: POWER_MODEL: ID=2 NAME=MODE_15W_6CORE
NVPM VERB: POWER_MODEL: ID=3 NAME=MODE_10W_2CORE
NVPM VERB: POWER_MODEL: ID=4 NAME=MODE_10W_4CORE
NVPM VERB: POWER_MODEL: ID=5 NAME=MODE_10W_DESKTOP
NVPM VERB: POWER_MODEL: ID=6 NAME=MODE_20W_2CORE
NVPM VERB: POWER_MODEL: ID=7 NAME=MODE_20W_4CORE
NVPM VERB: POWER_MODEL: ID=8 NAME=MODE_20W_6CORE

Activate max speed fan

> sudo vim /etc/nvfancontrol.conf

# --- Modify ----
	FAN_PROFILE quiet {
...
	}
	FAN_PROFILE full { # Add this
		#TEMP 	HYST	PWM	RPM
		0	8 	255 	6000
		108	0 	255 	6000
	}
	FAN_PROFILE cool {
...
	FAN_DEFAULT_CONTROL open_loop
	FAN_DEFAULT_PROFILE full # activate to full
	FAN_DEFAULT_GOVERNOR pid

# ---- Saved file ----
> sudo systemctl stop nvfancontrol
> sudo rm /var/lib/nvfancontrol/status
> sudo systemctl start nvfancontrol

Install TRTPose Model Api Install torch2trt for conversion form PyTorch model to TensorRT

> pip3 install tqdm cython pycocotools matplotlib # required packages
> git clone https://github.com/NVIDIA-AI-IOT/torch2trt
> cd torch2trt
> python3 setup.py install

Install TRTPose

> cd ../
> git clone https://github.com/NVIDIA-AI-IOT/trt_pose
> cd trt_pose/
> python3 setup.py install

with open('human_pose.json', 'r') as f:
    human_pose = json.load(f)

topology = trt_pose.coco.coco_category_to_topology(human_pose)
num_parts = len(human_pose['keypoints'])
num_links = len(human_pose['skeleton'])

MODEL_WEIGHTS = 'densenet121_baseline_att_256x256_B_epoch_160.pth'
OPTIMIZED_MODEL = 'densenet121_baseline_att_256x256_B_epoch_160_trt.pth'
model = trt_pose.models.densenet121_baseline_att(num_parts, 2 * num_links).cuda().eval()
WIDTH = 256
HEIGHT = 256

data = torch.zeros((1, 3, HEIGHT, WIDTH)).cuda()
if os.path.exists(OPTIMIZED_MODEL) == False:
    model.load_state_dict(torch.load(MODEL_WEIGHTS))
    model_trt = torch2trt.torch2trt(model, [data], fp16_mode=True, max_workspace_size=1<<25)
    torch.save(model_trt.state_dict(), OPTIMIZED_MODEL)

model_trt = TRTModule()
model_trt.load_state_dict(torch.load(OPTIMIZED_MODEL))

# ------- Results --------
- Converting ....
- Success

Clone Repo

> git clone https://github.com/tonible14012002/fall_detector
> cd fall_detector/
> python3 app.py --model trtpose --cam './sample-video'
# ---- Logs ----
Connected to server https://*************/

Contribution

...

tonible14012002 / fall_detector