Skip to content

Latest commit

 

History

History
172 lines (128 loc) · 7.4 KB

README.md

File metadata and controls

172 lines (128 loc) · 7.4 KB
Raw

detectnet-diamond.py

Jetson nano warns safty diamond-shaped symbols near the pedestrian crossing in Japan

In Japan, two diamond-shaped symbols known as "dia-mark" are printed on the road 50m before the pedestrian crossing to warn vehicle drivers that pedestrians crossing at crosswalks without signals.
Vehicles must stop if there are pedestrians attempting to cross. However, vehicles do not notice the diamond-shaped symbols and approach the pedestrian crossing at a dangerous speed.

detectnet-diamond.py uses a Jetson nano and a USB camera to detect the diamond-shaped symbols and outputs the rising pulse to GPIO38 where ISD1820 board is connected.

ISD1820 is a board that can record audio for max 10 seconds and plays back the audio if the rising pulse enters the PlayE terminal. If you record the appropriate voice such as barking dog , it will tell you audibly when Jetson nano recognizes the diamond-shaped symbols.

diamond

The first time it starts up, it takes about 10 minutes to run [TRT] TensorRT and starts recognition, but after the first time, it takes about 30 seconds to recognize the diamond-shaped symbols.

This program is derived from https://github.com/dusty-nv/jetson-inference/blob/master/python/examples/detectnet.py

Revision history

Low pass filter for confidence level is introduced to reduce false detection. rate=0.9 is default , means about 10 frames of confidence > 0.5 are necessary to detect a diamond.

Since sometimes gstreamer stops by "failed to retrieve next image" , detectnet-diamond.py dies and is restarted by detectnet-diamond.sh

[gstreamer] gstDecoder -- failed to retrieve next image buffer
Expecting value: line 1 column 1 (char 0)Expecting value: line 1 column 1 (char 0)Expecting value: line 1 column 1 (char 0)Extra data: line 1 column 4 (char 3)[gstreamer] gstreamer message qos ==> v4l2src0
Traceback (most recent call last):
  File "/home/xxxxxx/detectnet-diamond.py", line 150, in 
    img = input.Capture()
Exception: jetson.utils -- videoSource failed to capture image

Hard ware :

Jetson nano 2GB

logi C270n USB camera

ISD1820 voice recorder board

USB GPS : EPS32 with NEO6M GPS board or GPS module for M5stack.

ESP32 is programed to send NMEA to USB serial port. https://github.com/coniferconifer/ESP32-GPS-BTserial

pin Number 33 (gpio38) at J41 connector should be wired to ISD1820 PlayE(PE) pin
pin Number 2,4 can be used as VCC supply to ISD1820
any of one of pin Number 6,9,14,20,25,30,34,39 can be used as GND for ISD1820

https://jetsonhacks.com/nvidia-jetson-nano-j41-header-pinout/

If the system reacts incorrectly when the vehicle stops, it will continue to react for a long time, so the system uses the vehicle speed obtained from the GPS to recognize the vehicle at a speed of 20 km/h or more. (speedThresh) Also, since Jetson nano is not equipped with a battery-backed RTC, it is not possible to get a time that can be used for logging in environments without an Internet connection, so the time is obtained from GPS.

default TIMEZONE is JST

How to set up GPS and ntpd

$ sudo apt install gpsd
$ sudo apt install ntp
$ sudo systemctl stop systemd-timesyncd
$ sudo systemctl disable systemd-timesyncd

edit /etc/ntp.conf and add following lines to sync with GPS time.

server 127.127.28.0  minpoll 4 maxpoll 4 prefer
fudge 127.127.28.0 flag1 1 time1 0.0 refid GPS

reboot and confirm ntpd and gpsd are running

$ sudo reboot now
$ cgsp -s
$ ntpq -p
ntpq -p
     remote           refid      st t when poll reach   delay   offset  jitter
==============================================================================
*SHM(0)          .GPS.            0 l   14   16  377    0.000   31.569  33.813
 ntp.ubuntu.com  .POOL.          16 p    -   64    0    0.000    0.000   0.000

How to install related python module.

$ pip3.6 install Jetson.GPIO
$ pip3.6 install gps3
$ pip3.6 install pytz

setup User permissions

Ref: https://github.com/NVIDIA/jetson-gpio

$ sudo groupadd -f -r gpio
$ sudo usermod -a -G gpio <your_user_name>
$ sudo cp lib/python/Jetson/GPIO/99-gpio.rules /etc/udev/rules.d/
$ sudo udevadm control --reload-rules && sudo udevadm trigger

verify GPIO38 and GPIO200

$ echo 38 > /sys/class/gpio/export
$ echo out > /sys/class/gpio/gpio38/direction
$ echo 1 > /sys/class/gpio/gpio38/value
$ echo 200 > /sys/class/gpio/export
$ echo out > /sys/class/gpio/gpio200/direction
$ echo 1 > /sys/class/gpio/gpio200/value

run shell script

The first time it is started, it takes about 10 minutes for inference to begin. Since ./models/diamond/ssd-mobilenet.onnx is zipped to upload on github , please unzip before use.

./detectnet-diamond.py  --headless=true --camera=/dev/video0 --width=640 --height=480 --model=models/diamond/ssd-mobilenet.onnx --labels=models/diamond/labels.txt --input-blob=input_0 --output-cvg=scores --output-bbox=boxes

--headless=true should be removed when you run on Window system.

Example : serial output

time(sec) from start, date , lat , lon, speed(km/h) , number of diamond  ,confidence level

881.677 ,2023-03-17 23:48:08 , 33.318686 , 134.6852810833 , 0.0 , 1 , 0.56
Left,Right,CenterX  148.984375 , 326.25 , 237.6171875

This program is trained to detect diamond-shaped mark by transfer learning according to what is described in re-training SSD-mobilenet. Although the number of learned diamond-shaped marks is small, the program is able to detect diamond-shaped marks, but it sometimes mis-detects roadside objects,etc.

The C270n camera has a resolution of 1280x720 and a 60-degree angle of view but is treated as a 640x480 camera input. To avoid misrecognizing roadside objects, when the center position of the recognized diamond shape falls within 150 pixels of the left or right edge, it is excluded.

In Japanese:

横断歩道の前の菱形マーク(ダイヤマーク）の認識をJetson nanoで検出できるか試したものです。./models/diamond/ssd-mobilenet.onnx はわずかな枚数の映像で学習したので誤検出があります。そこで、confidenceレベルにフレーム毎のローパスフィルタを掛けることで誤検知を抑制しています。 rate=0.9だと10フレーム約0.3秒位confidenceが0.5以上ないとフィルターされたconfidenceE は0.5以上にならないので検出が0.3秒程度遅れますが、誤検知は減らせています。 カメラは視界の邪魔にならないようダッシュボードの上に耐震粘着テープで転倒しないように貼り付けています。

Ref: https://github.com/dusty-nv/jetson-inference/blob/master/docs/pytorch-ssd.md

 Re-training SSD-mobilenet

Ref: https://github.com/dusty-nv/jetson-inference/blob/master/docs/detectnet-console-2.md

 Locating Objects With MobileNet

Ref: https://github.com/dusty-nv/jetson-inference/blob/master/docs/detectnet-camera-2.md

 Running the Live Camera detection Demo

Ref: https://github.com/dusty-nv/jetson-inference/blob/master/docs/detectnet-example-2.md

 Coding Your Own Object Detection Program

Ref: https://www.youtube.com/watch?v=2XMkPW_sIGg

 Jetson AI Fundamentals - S3E5 - Training Object Detection Models

Ref: https://www.youtube.com/watch?v=wuyJwMVzeiA

 How to use an ISD1820 Voice Recorder Module