5.Pedestrian Detection - Cascade classifier

Cascade classifier

1. Casacde classifier : face detection

• Viola - Jones face detection
  • It could train positive image(face image) and negative image(normal image) and detect accuratly face area.
  • Different point with before methods is that Haar-like feature, robust classifier function based on AdaBoost, rapid operation speed through cascade method are used.

• Haar-like feature
  • Using the set of filter of rectangle form
  • Extracting the result value pixel value is minused from black rectangle area to whole white rectangle area.

• Cascade classifier
  • Although There is one or two face in normal image, other area is almost non-face area.
  • Conducting Mutistage inspection to skip Non-face are.

• CascadeClassifier code

cv2.CascadeClassifier.detectMultiScale(image, scaleFactor=None, minNeighbors=None, flags=None, minSize=None, maxSize=None) -> result

• image : input image
• scaleFactor : image shrinkage ratio. default is 1.1
• minNeighbors: Specifying how many neighbor rectangle is detected to set to fianal detection area.
• flags : not use
• minSize : min object size
• maxSize : max object size
• result : numpy.ndarray put in rectangle information of detected object such as (x,y,w,h)

2. Hisgogram of Oriented Gradients (Hog)

• Hog
  • Using oriented gradient of image as feature vector
  • It’s widely used to pedstrian detection mehod at CVPR conference in 2005

• The code Loading pre-learn clssifier coefficient for making HOG descriptor object or detecting pedestrian.

cv2.HOGDescriptor() -> <CascadeClassifier object>

cv2.HOGDescriptor_getDefaultPeopleDetector() -> retval

• retval : pre-trained feature vector

• Enrolling SVM classifier coefficient

cv2.HOGDescriptor.setSVMDetector(svmdetector) -> None

• svmdetector : coefficient for linear SVM classifier

• HOG multiscale object detection code

cv2.HOGDescriptor.detectMultiScale(img, hitThreshold=None, winStride=None, padding=None, scale=None, finalThreshold=None, useMeanshiftGrouping=None) -> foundLocations, foundWeights

• img : input image
• hitTreshold : Threshold for distance of between feature vector and SVM classifer plane
• winStride : Moving size of shall window
• padding : Padding size
• scale : Size ration of search window
• finalThreshold : Threshold for detection determination
• useMeanshiftGrouping : The method superimposed window combine
• foundLocations : Rectangle area information
• foundWeights : Confidence for rectangle area

• HOG pedestrian detection result example