Tuesday 26 September 2017

Motion Gesture Recognition within 200 lines using OpenCV +Python

Hello Readers. In this tutorial, I will be teaching you gesture recognition in OpenCV+Python using only Image Processing and no Machine Learning or any Neural Networks.

What is Gesture Recognition

Gesture recognition is the mathematical interpretation of a human motion by a computing device. There are many algorithms out on the Internet that gives a very good and accurate results on gesture recognition. But here we are not going see them. I developed a very simple and naive algorithm to recognize gestures that are made up of straight lines. Lets see....
Gesture Recognition

Give me the code already...

Ok. I got you. Here is the code https://github.com/EvilPort2/SimpleGestureRecognition.

Why no Machine Learning or Neural Net?

The answer to this question is very simple. I do not know much about them. Though I have some knowledge about machine learning, I have little to no knowledge about neural networks.

Requirements

  1. A computer with a good camera
  2. A yellow (though any other colour can be used) piece of paper to be worn in a finger (for image segmentation)
  3. OpenCV for Python 3
  4. PyAutoGui for Python 3
  5. Python 3
  6. Text Editor like Sublime Text 3 or Atom
  7. A little bit of knowledge in Maths

Steps we are taking

Since I am using only image processing for this project, I will be using only the direction of movement to determine the gesture.
  1. Take one frame at a time and convert it from RGB colour space to HSV colour space for better yellow colour segmentation.
  2. Use a mask for yellow colour.
  3. Bluring and thresholding the mask.
  4. If a yellow colour is found and it crosses a reasonable area threshold, we start to create a gesture.
  5. The direction of movement of the yellow cap is calculated by taking the difference between the old centre and the new centre of the yellow colour after every 5th iteration or frame.
  6. Take the directions and store in a list until the yellow cap disappears from the frame.
  7. Process the created direction list and the processed direction list is used to take a certain action like a keyboard shortcut.
 

Let's get our hands dirty...

gesture_action.py

Let us begin with all the important imports and a few global variables
import cv2 import numpy as np from collections import deque import pyautogui as gui from gesture_api import do_gesture_action cam = cv2.VideoCapture(0) # Camera Object yellow_lower = np.array([7, 96, 85]) # HSV yellow lower yellow_upper = np.array([255, 255, 255]) # HSV yellow upper screen_width, screen_height = gui.size() camx, camy = 360, 240 # Resize resolution buff = 128 line_pts = deque(maxlen = buff) # Create a deque data structure which store the present location of centre point of the yellow patch

The gesture_api is a different file that I created. do_gesture_action is a function in that file. The yellow_lower and yellow_upper can be determined by using this python program. So in your case, these values might be different in different lighting conditions. The easiest way to use it is to put the yellow paper in front of the camera and then slowly increasing the lower parameters(H_MIN, V_MIN, S_MIN) one by one and then slowly decreasing the upper parameters (H_MAX, V_MAX, S_MAX). When the adjusting has been done you will find that only the yellow paper will have a corresponding white patch and rest of the image will be dark. Now let's get into the main function and some of its local variables
def gesture_action(): centerx, centery = 0, 0 # Present location of the centre of the yellow patch old_centerx, old_centery = 0, 0 # Previous location of the centre of the yellow patch area1 = 0 # Area of the yellow patch c = 0 # Stores the number of yellow objects in the picture flag_do_gesture = 0 # If a gesture has been completed then this flag is 1 flag0 = True # Checks if a yellow object is present in the frame created_gesture_hand1 = [] # stores the direction of the movement

With that out of the way we can now extract each frame and do the operations as required. These are steps we will be doing
    1. Get a frame
    2. Flip and resize the image to 360*240 for faster processing
    3. Convert the frame from RGB colour space to HSV colour space
    4. Now we will be using the yellow colour mask to segment the yellow colour
    5. Because every camera has some flaws in them which introduces some error in the frame hence we need to reduce the noise in the image and the easiest way to do that is to heavily blur the frame.
    6. Now if we set the colour threshold to any colour which is above black then we can get the almost exact shape of the the yellow patch.
    7. Take the contour of the thresholded frame.
    8. Repeat the above steps for every frame

while True: _, img = cam.read() # Resize for faster processing. Flipping for better orientation img = cv2.flip(img, 1) img = cv2.resize(img, (camx, camy)) # Convert to HSV for better color segmentation imgHSV = cv2.cvtColor(img, cv2.COLOR_BGR2HSV) # Mask for yellow color mask = cv2.inRange(imgHSV, yellow_lower, yellow_upper) # Bluring to reduce noises blur = cv2.medianBlur(mask, 15) blur = cv2.GaussianBlur(blur , (5,5), 0) # Thresholding _,thresh = cv2.threshold(blur,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU) cv2.imshow("Thresh", thresh) _, contours, _ = cv2.findContours(thresh.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE)
After getting the contours we can have 2 cases-
  1. Number of contours is greater than zero then yellow colored objects are in the frame.
  2. Number of contours is zero then no yellow colored objects are in the frame.

Case 1- Yellow colored objects in the frame

  1. Assign 0 to flag_do_gesture.
  2. Take the contour that has the maximum area. Let us call this max_contour.
  3. Find a minimum area rectangle that surrounds the max_contour.
  4. Take the width and height of the rectangle.
  5. Find the area of the rectangle by width*height.
  6. If the area crosses a reasonable threshold then start making a gesture. I found the threshold experimenting with different values and in my case it was 450.
  7. If the area of the contour crosses the threshold then find the center of the yellow object.
  8. Draw a rectangular box around it.
  9. Draw a dot at the center.
  10. Append the center to the deque line_pts.
  11. Update the center after every 5th iteration or frame.
  12. At the 5th iteration take the difference between the old center (x1, y1) and new center (x2, y2). I have used diffx = (x2-x1) and diffy = (y2-y1).
  13. Hence values of diffx and diffy gives us the direction of movement.
  14. If the flag0 is False then append the direction to the created_gesture_hand1 list.
  15. Draw a line for all the points in line_pts
  16. Assign False to flag0.

Case 2- No yellow colored objects in the frame

  1. Empty the deque line_pts.
  2. Process the created_gesture_hand1 by removing the 'St' and the consecutive directions. Let us call it processed_gesture_hand1.
  3. If flag_do_gesture is 0 and processed_gesture_hand1 then take an action corresponding to a particular gesture.
  4. Assign 1 to flag_do_gesture. This avoids the gesture action to be run only once and not repeatedly.
  5. Empty created_hand_gesture.
  6. Assign True to flag0.
Enough said..... In a code it looks something like this
if len(contours) == 0: # Completion of a gesture line_pts = deque(maxlen = buff) # Empty the deque processed_gesture_hand1 = tuple(process_created_gesture(created_gesture_hand1)) if flag_do_gesture == 0: # flag_do_gesture to make sure that gesture runs only once and not repeatedly if processed_gesture_hand1 != (): do_gesture_action(processed_gesture_hand1) flag_do_gesture = 1 print(processed_gesture_hand1) # for debugging purposes created_gesture_hand1 = [] flag0 = True else: flag_do_gesture = 0 max_contour = max(contours, key = cv2.contourArea) rect1 = cv2.minAreaRect(max_contour) (w, h) = rect1[1] area1 = w*h if area1 > 450: center1 = list(rect1[0]) box = cv2.boxPoints(rect1) # to draw a rectangle box = np.int0(box) cv2.drawContours(img,[box],0,(0,0,255),2) centerx = center1[0] = int(center1[0]) # center of the rectangle centery = center1[1] = int(center1[1]) cv2.circle(img, (centerx, centery), 2, (0, 255, 0), 2) line_pts.appendleft(tuple(center1)) if c == 0: old_centerx = centerx old_centery = centery c += 1 diffx, diffy = 0, 0 if c > 5: # check after every 5 iteration the new center diffx = centerx - old_centerx diffy = centery - old_centery c = 0 if flag0 == False: # the difference between the old center and the new center determines the direction of the movement if abs(diffx) <=10 and abs(diffy) <= 10: created_gesture_hand1.append("St") elif diffx > 15 and abs(diffy) <= 15: created_gesture_hand1.append("E") elif diffx < -15 and abs(diffy) <= 15: created_gesture_hand1.append("W") elif abs(diffx) <= 15 and diffy < -15: created_gesture_hand1.append("N") elif abs(diffx) <= 15 and diffy > 15: created_gesture_hand1.append("S") elif diffx > 25 and diffy > 25: created_gesture_hand1.append("SE") elif diffx < -25 and diffy > 25: created_gesture_hand1.append("SW") elif diffx > 25 and diffy < -25: created_gesture_hand1.append("NE") elif diffx < -25 and diffy < -25: created_gesture_hand1.append("NW") for i in range(1, len(line_pts)): if line_pts[i - 1] is None or line_pts[i] is None: continue cv2.line(img, line_pts[i-1], line_pts[i], (0, 255, 0), 2) flag0 = False

The process_created_gesture function looks like this
def process_created_gesture(created_gesture): """ function to remove all the St direction and removes duplicate direction if they occur consecutively. """ if created_gesture != []: for i in range(created_gesture.count("St")): created_gesture.remove("St") for j in range(len(created_gesture)): for i in range(len(created_gesture) - 1): if created_gesture[i] == created_gesture[i+1]: created_gesture.remove(created_gesture[i+1]) break return created_gesture
So the whole file gesture_action.py looks like this.
import cv2 import numpy as np import pyautogui as gui from gesture_api import do_gesture_action from collections import deque cam = cv2.VideoCapture(0) yellow_lower = np.array([7, 96, 85]) # HSV yellow lower yellow_upper = np.array([255, 255, 255]) # HSV yellow upper screen_width, screen_height = gui.size() camx, camy = 480, 360 buff = 128 line_pts = deque(maxlen = buff) def process_created_gesture(created_gesture): """ function to remove all the St direction and removes duplicate direction if they occur consecutively. """ if created_gesture != []: for i in range(created_gesture.count("St")): created_gesture.remove("St") for j in range(len(created_gesture)): for i in range(len(created_gesture) - 1): if created_gesture[i] == created_gesture[i+1]: created_gesture.remove(created_gesture[i+1]) break return created_gesture def gesture_action(): centerx, centery = 0, 0 old_centerx, old_centery = 0, 0 area1 = 0 c = 0 flag_do_gesture = 0 flag0 = True created_gesture_hand1 = [] while True: _, img = cam.read() # Resize for faster processing. Flipping for better orientation img = cv2.flip(img, 1) img = cv2.resize(img, (camx, camy)) # Convert to HSV for better color segmentation imgHSV = cv2.cvtColor(img, cv2.COLOR_BGR2HSV) # Mask for yellow color mask = cv2.inRange(imgHSV, yellow_lower, yellow_upper) # Bluring to reduce noises blur = cv2.medianBlur(mask, 15) blur = cv2.GaussianBlur(blur , (5,5), 0) # Thresholding _,thresh = cv2.threshold(blur,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU) cv2.imshow("Thresh", thresh) _, contours, _ = cv2.findContours(thresh.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE) w, h = 0, 0 if len(contours) == 0: # Completion of a gesture line_pts = deque(maxlen = buff) # Empty the deque processed_gesture_hand1 = tuple(process_created_gesture(created_gesture_hand1)) if flag_do_gesture == 0: # flag_do_gesture to make sure that gesture runs only once and not repeatedly if processed_gesture_hand1 != (): do_gesture_action(processed_gesture_hand1) flag_do_gesture = 1 print(processed_gesture_hand1) # for debugging purposes created_gesture_hand1 = [] flag0 = True else: flag_do_gesture = 0 max_contour = max(contours, key = cv2.contourArea) rect1 = cv2.minAreaRect(max_contour) (w, h) = rect1[1] area1 = w*h if area1 > 450: center1 = list(rect1[0]) box = cv2.boxPoints(rect1) # to draw a rectangle box = np.int0(box) cv2.drawContours(img,[box],0,(0,0,255),2) centerx = center1[0] = int(center1[0]) # center of the rectangle centery = center1[1] = int(center1[1]) cv2.circle(img, (centerx, centery), 2, (0, 255, 0), 2) line_pts.appendleft(tuple(center1)) if c == 0: old_centerx = centerx old_centery = centery c += 1 diffx, diffy = 0, 0 if c > 5: # check after every 5 iteration the new center diffx = centerx - old_centerx diffy = centery - old_centery c = 0 if flag0 == False: # the difference between the old center and the new center determines the direction of the movement if abs(diffx) <=10 and abs(diffy) <= 10: created_gesture_hand1.append("St") elif diffx > 15 and abs(diffy) <= 15: created_gesture_hand1.append("E") elif diffx < -15 and abs(diffy) <= 15: created_gesture_hand1.append("W") elif abs(diffx) <= 15 and diffy < -15: created_gesture_hand1.append("N") elif abs(diffx) <= 15 and diffy > 15: created_gesture_hand1.append("S") elif diffx > 25 and diffy > 25: created_gesture_hand1.append("SE") elif diffx < -25 and diffy > 25: created_gesture_hand1.append("SW") elif diffx > 25 and diffy < -25: created_gesture_hand1.append("NE") elif diffx < -25 and diffy < -25: created_gesture_hand1.append("NW") for i in range(1, len(line_pts)): if line_pts[i - 1] is None or line_pts[i] is None: continue cv2.line(img, line_pts[i-1], line_pts[i], (0, 255, 0), 2) flag0 = False cv2.imshow("IMG", img) if cv2.waitKey(1) == ord('q'): break cv2.destroyAllWindows() cam.release() gesture_action()

gesture_api.py

This file contains nothing but the gesture directions and the keyboard shortcut that it needs to emulate. So a square can be made using directions like (North, West, South, East). Now let's say that when a square is made we need to emulate the keyboard shortcut winkey (For Windows) or altleft+f1 (For KDE) and so on. We can have 2 cases for the keyboard shortcut emulation.
  • Only one key press needs to be emulated e.g winkey
  • More than one key press needs to be emulated e.g winkey + l, alt + f4 etc.
For the first case, we need to just press the key. For the second case, we need to hold all the keys except the last key, press the last key and then un-hold the keys. In code this can be accomplished by-
import pyautogui as gui import os GEST_START = ("N", "E", "S", "W") GEST_CLOSE = ("SE", "N", "SW") GEST_COPY = ("W", "S", "E") GEST_PASTE = ("SE", "NE") GEST_CUT = ("SW", "N", "SE") GEST_ALT_TAB = ("SE", "SW") GEST_ALT_SHIFT_TAB = ("SW", "SE") GEST_MAXIMISE = ("N",) GEST_MINIMISE = ("S",) GEST_LOCK = ("S", "E") GEST_TASK_MANAGER = ("E", "W", "S") GEST_NEW_FILE = ("N", "SE", "N") GEST_SELECT_ALL = ("NE", "SE", "NW", "W") # Gesture set containing the directions and the key press actions GESTURES = {GEST_CUT: ('ctrlleft', 'x'), GEST_CLOSE: ('altleft', 'f4'), GEST_ALT_SHIFT_TAB: ('altleft', 'shiftleft', 'tab'), GEST_PASTE: ('ctrlleft', 'v'), GEST_ALT_TAB: ('altleft', 'tab'), GEST_COPY: ('ctrlleft', 'c'), GEST_NEW_FILE: ('ctrlleft', 'n'), GEST_SELECT_ALL: ('ctrlleft', 'a')} # Windows PCs if os.name == 'nt': GESTURES[GEST_START] = ('winleft',) GESTURES[GEST_LOCK] = ('winleft', 'l') GESTURES[GEST_TASK_MANAGER] = ('ctrlleft', 'shiftleft', 'esc') # Linux using KDE else: GESTURES[GEST_START] = ('altleft', 'f1') GESTURES[GEST_LOCK] = ('ctrlleft', 'altleft', 'l') GESTURES[GEST_TASK_MANAGER] = ('ctrlleft', 'esc') def do_gesture_action(gesture): if gesture in GESTURES.keys(): keys = list(GESTURES[gesture]) last_key = keys.pop() # get the last key press if len(keys) >= 1: # case 2 for key in keys: # hold all the keys except the last key gui.keyDown(key) gui.press(last_key) # press the last key. for case 1 the last key and the first key are the same if len(keys) >= 1: keys.reverse() # un-holding the keys for key in keys: gui.keyUp(key)
 

Conclusion

Yes. And that's about it. Using only 2 files and only image processing we have successfully implemented a very simple and naive gesture recognition system. That too happened within only 200 lines of code. Get the full code here. You can find me on-
Bye.....

2 comments:

  1. hi
    i am using pycharm
    its showing error for gesture_api
    can you please give a solution?

    ReplyDelete

Gesture driven Virtual Keyboard using OpenCV + Python

Hello Readers, long time no see. In this tutorial, I will be teaching you how to create a gesture driven Virtual Keyboard with OpenCV and P...