# -*- coding: utf-8 -*- """ Created on Mon Mar 4 18:28:47 2024 @author: ym """ import cv2 import numpy as np from scipy.spatial.distance import cdist from sklearn.decomposition import PCA from .dotracks import MoveState, Track from tools.config import cfg class backTrack(Track): # boxes: [x1, y1, x2, y2, track_id, score, cls, frame_index, box_index] # 0, 1, 2, 3, 4, 5, 6, 7, 8 def __init__(self, boxes, imgshape=(1024, 1280)): super().__init__(boxes, imgshape) '''(cent_x, cent_y, mean_w, mean_h)''' '''5个关键点（中心点、左上点、右上点、左下点、右下点 ）轨迹特征''' self.compute_cornpts_feats() '''该函数依赖项: self.cornpoints''' self.isCornpoint = self.isimgborder() '''该函数依赖项: self.cornpoints、self.isCornpoint，定义 10 个轨迹特征： self.trajectory, self.trajmin, self.trajlens, self.trajdist self.feature = [trajlen_min, trajlen_max, trajdist_min, trajdist_max, trajlen_rate, trajdist_rate] ''' self.trajfeature() '''静止点帧索引''' self.static_index = self.compute_static_fids() '''运动点帧索引(运动帧两端的静止帧索引)''' self.moving_index = self.compute_moving_fids() '''该函数依赖项: self.cornpoints，定义 4 个商品位置变量： self.Cent_isIncart, self.LB_isIncart, self.RB_isIncart self.posState = self.Cent_isIncart+self.LB_isIncart+self.RB_isIncart''' self.PositionState() '''self.feature_ious = (incart_iou, outcart_iou, cartboarder_iou, maxbox_iou, minbox_iou) self.incartrates = incartrates''' self.compute_ious_feat() # self.PCA() # ============================================================================= # def compute_cornpoints(self): # ''' # cornpoints 共10项，分别是个点的坐标值（x, y） # (center, top_left, top_right, bottom_left, bottom_right) # ''' # boxes = self.boxes # cornpoints = np.zeros((self.frnum, 10)) # cornpoints[:,0], cornpoints[:,1] = boxes[:, 0], boxes[:, 1] # cornpoints[:,2], cornpoints[:,3] = boxes[:, 0] - boxes[:, 2]/2, boxes[:, 1] - boxes[:, 3]/2 # cornpoints[:,4], cornpoints[:,5] = boxes[:, 0] + boxes[:, 2]/2, boxes[:, 1] - boxes[:, 3]/2 # cornpoints[:,6], cornpoints[:,7] = boxes[:, 0] - boxes[:, 2]/2, boxes[:, 1] + boxes[:, 3]/2 # cornpoints[:,8], cornpoints[:,9] = boxes[:, 0] + boxes[:, 2]/2, boxes[:, 1] + boxes[:, 3]/2 # # self.cornpoints = cornpoints # ============================================================================= def isimgborder(self, BoundPixel=10, BoundThresh=0.3): x1, y1 = self.cornpoints[:,2], self.cornpoints[:,3], x2, y2 = self.cornpoints[:,8], self.cornpoints[:,9] cont1 = sum(abs(x1) BoundThresh cont2 = sum(abs(y1) BoundThresh cont3 = sum(abs(x2-self.imgshape[0]) BoundThresh cont4 = sum(abs(y2-self.imgshape[1]) BoundThresh cont = cont1 or cont2 or cont3 or cont4 isCornpoint = False if cont: isCornpoint = True return isCornpoint def PositionState(self, camerType="back"): ''' camerType: back, 后置摄像头 front, 前置摄像头 ''' if camerType=="front": incart = cv2.imread(cfg.incart, cv2.IMREAD_GRAYSCALE) else: incart = cv2.imread(cfg.incart_ftmp, cv2.IMREAD_GRAYSCALE) xc, yc = self.cornpoints[:,0].clip(0,self.imgshape[0]-1).astype(np.int64), self.cornpoints[:,1].clip(0,self.imgshape[1]-1).astype(np.int64) x1, y1 = self.cornpoints[:,6].clip(0,self.imgshape[0]-1).astype(np.int64), self.cornpoints[:,7].clip(0,self.imgshape[1]-1).astype(np.int64) x2, y2 = self.cornpoints[:,8].clip(0,self.imgshape[0]-1).astype(np.int64), self.cornpoints[:,9].clip(0,self.imgshape[1]-1).astype(np.int64) # print(self.tid) Cent_inCartnum = np.count_nonzero(incart[(yc, xc)]) LB_inCartnum = np.count_nonzero(incart[(y1, x1)]) RB_inCartnum = np.count_nonzero(incart[(y2, x2)]) self.Cent_isIncart = False self.LB_isIncart = False self.RB_isIncart = False if Cent_inCartnum: self.Cent_isIncart = True if LB_inCartnum: self.LB_isIncart = True if RB_inCartnum: self.RB_isIncart = True self.posState = self.Cent_isIncart+self.LB_isIncart+self.RB_isIncart def PCA(self): self.pca = PCA() X = self.cornpoints[:, 0:2] self.pca.fit(X) def compute_ious_feat(self): '''输出： self.feature_ious = (incart_iou, outcart_iou, cartboarder_iou, maxbox_iou, minbox_iou) self.incartrates = incartrates， 其中： boxes流：track中所有boxes形成的轨迹图，可分为三部分：incart, outcart, cartboarder incart_iou, outcart_iou, cartboarder_iou：各部分和 boxes流的 iou。 incart_iou = 0，track在购物车外， outcart_iou = 0，track在购物车内，也可能是通过左下角、右下角置入购物车， maxbox_iou, minbox_iou：track中最大、最小 box 和boxes流的iou，二者差值越小，越接近 1，表明track的运动型越小。 incartrates: 各box和incart的iou时序，由小变大，反应的是置入过程，由大变小，反应的是取出过程 ''' incart = cv2.imread(cfg.incart, cv2.IMREAD_GRAYSCALE) outcart = cv2.imread(cfg.outcart, cv2.IMREAD_GRAYSCALE) cartboarder = cv2.imread(cfg.cartboarder, cv2.IMREAD_GRAYSCALE) incartrates = [] temp = np.zeros(incart.shape, np.uint8) maxarea, minarea = 0, self.imgshape[0]*self.imgshape[1] for i in range(self.frnum): # x, y, w, h = self.boxes[i, 0:4] x = (self.boxes[i, 2] + self.boxes[i, 0]) / 2 w = (self.boxes[i, 2] - self.boxes[i, 0]) / 2 y = (self.boxes[i, 3] + self.boxes[i, 1]) / 2 h = (self.boxes[i, 3] - self.boxes[i, 1]) / 2 if w*h > maxarea: maxarea = w*h if w*h < minarea: minarea = w*h cv2.rectangle(temp, (int(x-w/2), int(y-h/2)), (int(x+w/2), int(y+h/2)), 255, cv2.FILLED) temp1 = np.zeros(incart.shape, np.uint8) cv2.rectangle(temp1, (int(x-w/2), int(y-h/2)), (int(x+w/2), int(y+h/2)), 255, cv2.FILLED) temp2 = cv2.bitwise_and(incart, temp1) inrate = cv2.countNonZero(temp1)/(w*h) incartrates.append(inrate) isincart = cv2.bitwise_and(incart, temp) isoutcart = cv2.bitwise_and(outcart, temp) iscartboarder = cv2.bitwise_and(cartboarder, temp) num_temp = cv2.countNonZero(temp) num_incart = cv2.countNonZero(isincart) num_outcart = cv2.countNonZero(isoutcart) num_cartboarder = cv2.countNonZero(iscartboarder) incart_iou = num_incart/num_temp outcart_iou = num_outcart/num_temp cartboarder_iou = num_cartboarder/num_temp maxbox_iou = maxarea/num_temp minbox_iou = minarea/num_temp self.feature_ious = (incart_iou, outcart_iou, cartboarder_iou, maxbox_iou, minbox_iou) self.incartrates = incartrates def compute_static_fids(self, thresh1 = 12, thresh2 = 3): ''' 计算 track 的轨迹中相对处于静止状态的轨迹点的（start_frame_id, end_frame_id） thresh1: 相邻两帧目标中心点是否静止的的阈值，以像素为单位， thresh2: 连续捕捉到目标处于静止状态的帧数 ''' BoundPixel = 8 x1, y1 = self.cornpoints[:,2], self.cornpoints[:,3], x2, y2 = self.cornpoints[:,8], self.cornpoints[:,9] cont1 = sum(abs(x1) 3 # cont2 = sum(abs(y1) 3 cont3 = sum(abs(x2-self.imgshape[0]) 3 # cont4 = sum(abs(y2-self.imgshape[1]) 3 cont = not(cont1 or cont3) ## ============== 下一步，启用中心点，选择具有最小运动幅度的角点作为参考点 static_index = [] if self.frnum>=2 and cont: x1 = self.boxes[1:,7] x2 = [i for i in range(int(min(x1)), int(max(x1)+1))] dist_adjc = np.interp(x2, x1, self.trajmin) # dist_adjc = self.trajmin static_thresh = (dist_adjc < thresh1)[:, None].astype(np.uint8) static_cnts, _ = cv2.findContours(static_thresh, cv2.RETR_LIST, cv2.CHAIN_APPROX_NONE) for cnt in static_cnts: _, start, _, num = cv2.boundingRect(cnt) end = start + num if num <= thresh2: continue static_index.append((start, end)) # ============================================================================= # '''========= 输出帧id，不太合适 =========''' # static_fids = [] # for i in range(len(static_index)): # i1, i2 = static_index[i] # fid1, fid2 = boxes[i1, 7], boxes[i2, 7] # static_fids.append([fid1, fid2]) # static_fids = np.array(static_fids) # ============================================================================= static_index = np.array(static_index) if static_index.size: indx = np.argsort(static_index[:, 0]) static_index = static_index[indx] return static_index def compute_moving_fids(self, thresh1 = 12, thresh2 = 3): ''' 计算 track 的轨迹中运动轨迹点的（start_frame_id, end_frame_id） thresh1: 相邻两帧目标中心点是否运动的阈值，以像素为单位， thresh2: 连续捕捉到目标连续运动的帧数 目标： 1. 计算轨迹方向 2. 计算和手部运动的关联性 ''' moving_index = [] if self.frnum>=2: x1 = self.boxes[1:,7] x2 = [i for i in range(int(min(x1)), int(max(x1)+1))] dist_adjc = np.interp(x2, x1, self.trajmin) moving_thresh = (dist_adjc >= thresh1)[:, None].astype(np.uint8) moving_cnts, _ = cv2.findContours(moving_thresh, cv2.RETR_LIST, cv2.CHAIN_APPROX_NONE) for cnt in moving_cnts: _, start, _, num = cv2.boundingRect(cnt) if num < thresh2: continue end = start + num moving_index.append((start, end)) # ============================================================================= # '''========= 输出帧id，不太合适 =========''' # moving_fids = [] # for i in range(len(moving_index)): # i1, i2 = moving_index[i] # fid1, fid2 = boxes[i1, 7], boxes[i2, 7] # moving_fids.append([fid1, fid2]) # moving_fids = np.array(moving_fids) # ============================================================================= moving_index = np.array(moving_index) if moving_index.size: indx = np.argsort(moving_index[:, 0]) moving_index = moving_index[indx] return moving_index def compute_distance(self): pass def move_start_fid(self): pass def move_end_fid(self): pass