Flow chart

CangHai • Posted by vany5921

Published March 30, 2020 © LGPL

M5StickV deep learning for WeChat Jump game

Thank @沧海 for his contribution We choose wechat jump, a classic little game, as a carrier of target detection test.

AdvancedFull instructions provided3 hours1,127

M5StickV deep learning for WeChat Jump game

Things used in this project

Hardware components

M5Stack M5StickV

Touch screen Clicker

Software apps and online services

M5Stack V-Training

Story

M5stickv and the subsequent unit-v are the visual sensor modules of m5stack.com (眀栈科技) based on Kendryte K210, which can perform high-speed convolution neural network calculation under ultra-low power consumption. Use scenarios such as target detection and image classification tasks based on convolutional neural network, face detection and face recognition, multi class object detection and recognition, etc.So we choose the classic game of wechat jump as a carrier of target detection test.

1. Object Detection

In this concept, I have carried a passage from "YOLO principle and Implementation of object detection" by Xiaobaijiang

object detection is a more practical and challenging computer vision task, which can be seen as the combination of image classification and location. Given a picture, the object detection system should be able to identify the object of the picture and give its location. Because the number of object in the picture is uncertain and the exact location of the object should be given, object detection is more complex than classification task.

A practical application scenario of object detection is driverless. If an effective object detection system can be loaded on the driverless vehicle, the driverless vehicle will have eyes like the human, and can quickly detect the pedestrians and vehicles in front, so as to make real-time decision.

Well, let's apply this phrase to the scene of "wechat jump", that is, we use m5stickv to classify the puppets in the "jump" screen and the platform (identify which is the puppet and which is the platform), and then determine the specific location between the puppet and the platform to be reached (obtain the center point coordinates). Because the number of jumps in the field of vision is uncertain (there are already skipped, there is the next target jump), we need to get the exact location of the target. Obviously, target detection is more complex than a single classification task.

2, yolo3

As the author will explain, Yolo algorithm, its full name is you only look once: unified, real time object detection. The name of the algorithm is very good, basically summarizing the characteristics of Yolo algorithm:

You only look once means that only one CNN operation is needed. Unified means that this is a unified framework that provides end-to-end prediction, while real time means that Yolo algorithm is fast.

M5StickV (Unit-V) Design Guru smile more grounded: K210 is mainly to provide a simple method for users who do not understand AI to use, inside running is the neural network structure of Google's MobilenetV1, plus the detection structure of YoloV3. The core of Yolo is you only look once, which directly gets the position and class of the target through a picture. A single operation can predict hundreds of targets at most, and the detection time will not increase with the number of targets.

Knowing these two concepts, we as players know that m5stickv (unit-v) can provide us with target detection services, so that we can identify the puppet and the platform and get their exact positions in the wechat hop.

Flow chart

Step 1

shoot more than 100 samples. It is recommended that the shooting should be consistent with the angle environment you actually use (cell phone, light conditions, shooting angle, etc.) as close as possible.According to Hanxiao's suggestion, I modified a simple automatic shooting program. I took a picture every 3S and saved it in the / train directory. It needs to run under the firmware m5stickv ﹣ firmware ﹣ 1022 ﹣ beta.kfpkg. Attached.

Step 2

mark the object.

Step3

upload zip file to http://v-training.m5stack.com/

Step 4

after the server training, receive the feedback model, firmware and boot.py file

import audio
import gc
import image
import lcd
import sensor
import sys
import time
import uos
import os
#import KPU as kpu
from fpioa_manager import *
from machine import I2C
from Maix import I2S, GPIO
 
#
# initialize
#
lcd.init()
lcd.rotation(2)
 
fm.register(board_info.SPK_SD, fm.fpioa.GPIO0)
spk_sd=GPIO(GPIO.GPIO0, GPIO.OUT)
spk_sd.value(1) #Enable the SPK output
 
fm.register(board_info.SPK_DIN,fm.fpioa.I2S0_OUT_D1)
fm.register(board_info.SPK_BCLK,fm.fpioa.I2S0_SCLK)
fm.register(board_info.SPK_LRCLK,fm.fpioa.I2S0_WS)
 
wav_dev = I2S(I2S.DEVICE_0)
 
#fm.register(board_info.BUTTON_A, fm.fpioa.GPIO1)
#but_a=GPIO(GPIO.GPIO1, GPIO.IN, GPIO.PULL_UP) #PULL_UP is required here!
 
#fm.register(board_info.BUTTON_B, fm.fpioa.GPIO2)
#but_b = GPIO(GPIO.GPIO2, GPIO.IN, GPIO.PULL_UP) #PULL_UP is required here!
 
currentImage=0
 
def play_sound(filename):
    try:
        player = audio.Audio(path = filename)
        player.volume(20)
        wav_info = player.play_process(wav_dev)
        wav_dev.channel_config(wav_dev.CHANNEL_1, I2S.TRANSMITTER,resolution = I2S.RESOLUTION_16_BIT, align_mode = I2S.STANDARD_MODE)
        wav_dev.set_sample_rate(wav_info[1])
        spk_sd.value(1)
        while True:
            ret = player.play()
            if ret == None:
                break
            elif ret==0:
                break
        player.finish()
        spk_sd.value(0)
    except:
        pass
 
def initialize_camera():
    err_counter = 0
    while 1:
        try:
            sensor.reset() #Reset sensor may failed, let's try some times
            break
        except:
            err_counter = err_counter + 1
            if err_counter == 20:
                lcd.draw_string(lcd.width()//2-100,lcd.height()//2-4, "Error: Sensor Init Failed", lcd.WHITE, lcd.RED)
            time.sleep(0.1)
            continue
 
    sensor.set_pixformat(sensor.RGB565)
    sensor.set_framesize(sensor.QVGA) #QVGA=320x240
    sensor.run(1)
 
try:
    img = image.Image("/sd/startup.jpg")
    lcd.display(img)
except:
    lcd.draw_string(lcd.width()//2-100,lcd.height()//2-4, "Error: Cannot find start.jpg", lcd.WHITE, lcd.RED)
 
time.sleep(2)
 
initialize_camera()
 
#currentDirectory = 1
 
if "sd" not in os.listdir("/"):
    lcd.draw_string(lcd.width()//2-96,lcd.height()//2-4, "Error: Cannot read SD Card", lcd.WHITE, lcd.RED)
 
try:
    os.mkdir("/sd/train")
except Exception as e:
    pass
 
isButtonPressedA = 0
 
 
try:
    while(True):
        img = sensor.snapshot()
        disp_img=img.copy()
        disp_img.draw_rectangle(0,60,320,1,color=(0,144,255),thickness=10)
        #disp_img.draw_string(50,55,"Train:%03d/35   Class:%02d/10"%(currentImage,currentDirectory),color=(255,255,255),scale=1)
        lcd.display(disp_img)
 
        #if but_a.value() == 0 and isButtonPressedA == 0:
        img.save("/sd/train/" + str(currentImage) + ".jpg", quality=95)
        play_sound("/sd/kacha.wav")
        time.sleep(3)  #每隔三秒拍摄一张
        currentImage = currentImage + 1
        #    isButtonPressedA = 1
 
        #if but_a.value() == 1:
        #    isButtonPressedA = 0
 
 
except KeyboardInterrupt:
    pass

import image
import lcd
import sensor
import sys
import time
import KPU as kpu
from fpioa_manager import *
import math
import KPU as kpu
from Maix import GPIO
import utime
 
#使用grove的g35作为继电器执行pin，注册pin以及初始化
fm.register(35,fm.fpioa.GPIOHS0)
relay = GPIO(GPIO.GPIOHS0,GPIO.OUT)
 
relay.value(0)
#lcd初始化
lcd.init()
lcd.rotation(2)
#载入已训练模型
task=kpu.load("/sd/yolov3_mbnetv1_0.5_jumpAjump_100epoch_voc_v3.kmodel")
 
anchor = (0.33340788 * 16, 0.70065861 * 16, 0.18124964 * 16, 0.38986752 * 16, 0.08497349 * 16, 0.1527057 * 16)
 
a = kpu.init_yolo2(task, 0.05, 0.05, 3, anchor)
#注：第一个是兴趣框的精确度的阈值
#第二个参数，是目标的精确度的阈值，现在是0.05，调高了就会更可靠，但是可能识别率低
#调低了，容易识别到，但是也同时容易误识别。
print("Load Done.")
 
sensor.reset()
sensor.set_pixformat(sensor.RGB565)
sensor.set_framesize(sensor.QVGA)
sensor.set_windowing((320, 224))
sensor.run(1)
 
lcd.clear()
 
print("Init Done.")
 
counter = 1 #计数器
code_stake = [] #目标检测集合
minh = 1000 #目标跳台坐标
minh_y = 0
toy_x = 0 #人偶坐标
toy_y = 0
pressed_time = 0.0 #模拟手指的触屏时间
 
distance_factor = 0.25  #跳跃距离系数，此系数经过与实际手机屏幕校正后，据实际情况使用
 
while(True):
    img = sensor.snapshot()
    code = kpu.run_yolo2(task, img)
        #print("code....",code）
    if code:
        if counter < 5: #累计5次识别，将识别结果累加到code_stake[]当中
            code_stake = code_stake + code
            counter = counter  + 1
        else:
            counter = 0
            minh = 1000
            minh_y = 0
            toy_x = 0
            toy_y = 0
                        #注：i.rect()[0]   i.rect()[1]    i.rect()[2]    i.rect()[3]   分别是兴趣框的x,y,w,h
            for i in code_stake:
                if i.classid() == 0:  #识别到跳台,可能镜头里有多个跳台，我们需跳跃的是离摄像头最远的跳台
                    if minh > i.rect()[0] + i.rect()[2]//2:
                        minh = i.rect()[0] + i.rect()[2]//2  #计算跳台兴趣框中心点横坐标
                        minh_y = i.rect()[1] + i.rect()[3]//2#计算跳台兴趣框中心点纵坐标
                else: #识别到人偶
                    toy_x = i.rect()[0] + i.rect()[2]//2 #计算人偶兴趣框中心点横坐标
                    toy_y = i.rect()[1] + i.rect()[3]//2 #计算人偶兴趣框中心点纵坐标
            pressed_time = math.sqrt((minh - toy_x)**2 + (minh_y-toy_y)**2) / 100 * distance_factor #02-25更正，(minh_y-toy_y)
                        #pressed_time =人偶中心点至跳台中心点距离*跳跃距离系数
        img.draw_arrow(toy_x, toy_y, minh, minh_y)  #画人偶至跳台中心点连线，表明跳跃路径
        img.draw_string((toy_x+minh)//2, (minh_y+toy_y)//2, "%.03f" % (pressed_time), lcd.BLACK) #显示跳跃参数
 
        for i in code:
            if i.classid() == 0: #涉及跳台为红色标记
                c = lcd.RED
            else:
                c = lcd.GREEN #涉及人偶为绿色标记
            img.draw_circle(i.rect()[0] + i.rect()[2]//2, i.rect()[1] + i.rect()[3]//2, 3)#目标（跳台或人偶）中心点标记
            img.draw_rectangle(i.rect()[0], i.rect()[1], i.rect()[2], i.rect()[3], c) #标注目标（跳台或人偶）兴趣框
            lcd.display(img)   
 
        if counter == 0: #5次识别完毕
            relay.value(1) #继电器触发，模拟手指触屏动作
            utime.sleep_ms(int(pressed_time * 1000)) #经过pressed_time延时
            relay.value(0) #继电器接点返回，模拟抬起手指
            code_stake = [] #识别集清零
            utime.sleep_ms(2000)
 
 
    else:   
        lcd.display(img)
a = kpu.deinit(task)

Credits

CangHai

Posted by

vany5921

M5StickV deep learning for WeChat Jump game

Things used in this project

Hardware components

Software apps and online services

Story

Flow chart

Code

Autocapture program

M5StickV Controller

Credits

Comments

Embed the widget on your own site

M5StickV deep learning for WeChat Jump game

M5StickV deep learning for WeChat Jump game

Things used in this project

Hardware components

Software apps and online services

Story

Flow chart

Code

Autocapture program

M5StickV Controller

Credits

Comments

Related channels and tags