camshift achieves target tracking

注：The class for object tracking is not presentopencv-python库里,but with additional functionsopencv-contrib-python
安装：

pip install opencv-contrib-python==4.6.0.66 -i https://pypi.tuna.tsinghua.edu.cn/simple/

camshift Convert the image to a color probability distribution map using the color histogram model of the target,Initialize a search window size Small and location,And adaptively adjust the position and size of the search window according to the results obtained in the previous frame,Thereby, the center position of the target in the current image is located

实现步骤：
1 初始化搜索窗
2 反向投影：Convert the image to be less affected by lightingHSV模型.对HThe components are histograms representing the differencesHThe probability of occurrence of the value,Get the color probability lookup table.The color probability distribution map is obtained by replacing each pixel value of the image with the color occurrence probability
3 运行meanshiftThe algorithm obtains the new size and position of the search window.meanshiftIteratively finds the probability extremum to locate the target
4 重复2,3Step trace image

track superclass：
Image tracking images will inherit from this tracking superclass.This class implements the mouse to frame out the range of the image to be tracked,显示图像,显示CPS和RES值.

import cv2
import numpy as np 
import time

class TrackerBase(object):
    def __init__(self, window_name):
        self.window_name = window_name
        self.frame = None
        self.frame_width = None
        self.frame_height = None
        self.frame_size = None
        self.drag_start = None
        self.selection = None 
        self.track_box = None
        self.detect_box = None
        self.display_box = None
        self.marker_image = None
        self.processed_image = None
        self.display_image = None 
        self.target_center_x = None
    
    def onMouse(self, event, x, y, flags, params): 
        if self.frame is None:
            return
        if event == cv2.EVENT_LBUTTONDOWN and not self.drag_start:
            self.track_box = None 
            self.detect_box = None
            self.drag_start = (x, y) 
        if event == cv2.EVENT_LBUTTONUP: 
            self.drag_start = None
            self.detect_box = self.selection
        if self.drag_start: 
            xmin = max(0, min(x, self.drag_start[0]))
            ymin = max(0, min(y, self.drag_start[1]))
            xmax = min(self.frame_width, max(x, self.drag_start[0])) 
            ymax = min(self.frame_height, max(y, self.drag_start[1])) 
            self.selection = (xmin, ymin, xmax-xmin, ymax-ymin)
            
    def display_selection(self):
        if self.drag_start and self.is_rect_nonzero(self.selection):
            x, y, w, h = self.selection
            cv2.rectangle(self.marker_image, (x, y), (x + w, y + h), (0, 255, 255), 2)
            
    def is_rect_nonzero(self, rect):
        try:
            (_,_,w,h) = rect 
            return ((w>0)and(h>0)) 
        except: 
            try:
                ((_,_),(w,h),a) = rect 
                return (w > 0) and (h > 0) 
            except:
                return False
    def rgb_image_callback(self, data):
        frame = data
        if self.frame is None: 
            self.frame = frame.copy()
            self.marker_image = np.zeros_like(frame)
            self.frame_size = (frame.shape[1], frame.shape[0])
            self.frame_width, self.frame_height = self.frame_size 
            cv2.imshow(self.window_name, self.frame) 
            cv2.setMouseCallback(self.window_name,self.onMouse)
            cv2.waitKey(3) 
        else:
            self.frame = frame.copy() 
            self.marker_image = np.zeros_like(frame) 
            processed_image = self.process_image(frame) 
            self.processed_image = processed_image.copy() 
            self.display_selection()
            self.display_image = cv2.bitwise_or(self.processed_image, self.marker_image)
            
            if self.track_box is not None and self.is_rect_nonzero(self.track_box): 
                tx, ty, tw, th = self.track_box 
                cv2.rectangle(self.display_image, (tx, ty), (tx+tw, ty+th), (0, 0, 0), 2)
            elif self.detect_box is not None and self.is_rect_nonzero(self.detect_box):
                dx, dy, dw, dh = self.detect_box 
                cv2.rectangle(self.display_image, (dx, dy), (dx+dw, dy+dh), (255, 50, 50), 2)
            
            cv2.imshow(self.window_name, self.display_image) 
            cv2.waitKey(3)
            
    
    def process_image(self, frame):
        return frame
    
if __name__=="__main__": 
    cap = cv2.VideoCapture(0)
    trackerbase = TrackerBase('base')
    while True: 
        ret, frame = cap.read()
        x, y = frame.shape[0:2] 
        small_frame = cv2.resize(frame, (int(y/2), int(x/2)))
        trackerbase.rgb_image_callback(small_frame)
        if cv2.waitKey(1) & 0xFF == ord('q'): 
            break
    cap.release() 
    cv2.destroyAllWindows()

1 def onMouse(self, event, x, y, flags, params):
onMouseThe class is the default callback function for mouse events,在后面cv2.setMouseCallback(self.window_name,self.onMouse)使用到

参数：event：鼠标动作
x,y：When the mouse generates this eventx和y坐标
flags：cv2_EVENT_FLAG_* (MouseEventFlags)类型的变量
param：Custom passed to setMouseCallback 函数调用的参数

Put when left clickedx,y保存在变量drag_startas the initial position of the mouse,The new one will be placed at any time during mouse movementx,y传入回调函数,更新selection的值.Save the selected frame as when the left button is updetect_box

For more information about the mouse callback function, please refer to：http://t.csdn.cn/DuSMv

2 def display_selection(self):
Outline the selected area with a rectangle

3 def is_rect_nonzero(self, rect):
Determine whether the rectangle size is 0

4 def rgb_image_callback(self, data):
核心函数,有以下功能：
1 如没有frame时复制data（i.e. the incoming image）,Keep some parameters the same as the incoming image,并显示图像.Here, the mouse callback function is called to get the border

2 when there is an image,Run the image processing programself.process_image（The program is overridden in a subclass）,显示矩形框.显示track_box和detect_box

5 def process_image(self, frame):
The program currently just returns the original image,Subclasses should rewrite the program to implement the core function of tracking

camshift算法实现：

import cv2
import numpy as np 
from tracker_base import TrackerBase

class Camshift(TrackerBase):
    def __init__(self, window_name): 
        super(Camshift, self).__init__(window_name) 
        self.detect_box = None 
        self.track_box = None
        
    def process_image(self, frame):
        try: 
            if self.detect_box is None:
                return frame
            src = frame.copy()
            if self.track_box is None or not self.is_rect_nonzero(self.track_box): 
                self.track_box = self.detect_box
                x,y,w,h = self.track_box
                self.roi = cv2.cvtColor(frame[y:y+h, x:x+w], cv2.COLOR_BGR2HSV)
                roi_hist = cv2.calcHist([self.roi], [0], None, [16], [0, 180])
                self.roi_hist = cv2.normalize(roi_hist, roi_hist, 0, 255, cv2.NORM_MINMAX)
                self.term_crit = (cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, 10, 1)
            else:
                hsv = cv2.cvtColor(frame,cv2.COLOR_BGR2HSV)
                back_project = cv2.calcBackProject([hsv],[0],self.roi_hist,[0,180],1)
                ret, self.track_box = cv2.CamShift(back_project, self.track_box, self.term_crit)
                pts = cv2.boxPoints(ret)
                pts = np.int0(pts)
                cv2.polylines(frame,[pts],True,255,1)
        except:
            pass
        return frame

if __name__ == '__main__':
    cap = cv2.VideoCapture(0) 
    camshift = Camshift('camshift')
    while True:
        ret, frame = cap.read()
        x, y = frame.shape[0:2]
        small_frame = cv2.resize(frame, (int(y/2), int(x/2)))
        camshift.rgb_image_callback(small_frame) 
        if cv2.waitKey(1) & 0xFF == ord('q'): 
            break
    cap.release() 
    cv2.destroyAllWindows()

1

 try: 
            if self.detect_box is None:
                return frame

If there is no calibration tracking area, return to the original image directly

2

 if self.track_box is None or not self.is_rect_nonzero(self.track_box): 
                self.track_box = self.detect_box
                x,y,w,h = self.track_box
                self.roi = cv2.cvtColor(frame[y:y+h, x:x+w], cv2.COLOR_BGR2HSV)
                roi_hist = cv2.calcHist([self.roi], [0], None, [16], [0, 180])
                self.roi_hist = cv2.normalize(roi_hist, roi_hist, 0, 255, cv2.NORM_MINMAX)
                self.term_crit = (cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, 10, 1)

track_boxPut when it doesn't existtrack_box重置为detect_box（That is, the delineated tracking area）,并计算detect_boxColor probability histogram,This part is used for initialization

3

hsv = cv2.cvtColor(frame,cv2.COLOR_BGR2HSV)
                back_project = cv2.calcBackProject([hsv],[0],self.roi_hist,[0,180],1)
                ret, self.track_box = cv2.CamShift(back_project, self.track_box, self.term_crit)
                pts = cv2.boxPoints(ret)
                pts = np.int0(pts)
                cv2.polylines(frame,[pts],True,255,1)

调用CamshiftThe library recognizes images