记录下以前用opencv训练物体识别的几个命令及相关参数

# ls ./neg/*.* > neg.txt

# opencv_createsamples -vec ./pos.vec -info pos.txt -bg ./neg.txt -w 100 -h 100 -num 338

# opencv_traincascade -data ./data -vec ./pos.vec -bg ./neg.txt -numPos 200 -minHitRate 0.95 -numNeg 208 -numStages 16 -precalcValbufSize 2048 -precalcdxBufSize 2048 -w 100 -h 100

# nohup opencv_traincascade -data ./data -vec ./pos.vec -bg ./neg.txt -numPos 200 -minHitRate 0.95 -numNeg 208 -numStages 16 -precalcValbufSize 2048 -precalcdxBufSize 2048 -w 100 -h 100 > log.txt 2>/dev/null &

# BACKGROUNG PROCESSING TIME 是负样本切割时间 很长很长

# N 为训练层数 %SMP 样本占总样本个数 ST.THR 阈值，HR 击中率， FA  虚警 EXP.ERR 经验错误率

# opencv_createsamples -vec ./pos.vec -info pos.txt -bg ./neg.txt -w 24 -h 24 -num 338 -maxxangle 0.1 -maxyangle 0.1 -maxzangle 0.1

# opencv_traincascade -mode ALL -data ./data -vec ./pos.vec -bg ./neg.txt -numPos 200 -minHitRate 0.95 -numNeg 208 -numStages 16 -precalcValbufSize 2048 -precalcdxBufSize 2048 -w 24 -h 24

# nohup opencv_traincascade -mode ALL -data ./data -vec ./pos.vec -bg ./neg.txt -numPos 200 -minHitRate 0.95 -numNeg 208 -numStages 16 -precalcValbufSize 2048 -precalcdxBufSize 2048 -w 24 -h 24 > log.txt 2>/dev/null &

# http://www.cnblogs.com/whowhoha/p/5623977.html

# http://answers.opencv.org/question/4368/traincascade-error-bad-argument-can-not-get-new-positive-sample-the-most-possible-reason-is-insufficient-count-of-samples-in-given-vec-file/#4474

# numPos+（numStages-1）*numPos*（1-minHitRate）= 338

训练得到模型，测试脚本

detector = cv2.CascadeClassifier("./model.xml")

img = cv2.resize(img, (self.scalewidth, self.scaleheight))
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

rects = detector.detectMultiScale(gray, scaleFactor=1.1, minNeighbors=3, minSize=(50, 50))

加上dlib的correlation_tracker就可以实现对简单物体的实时跟踪。