yolov8-pose姿势估计，站立识别：动作检测-站姿-接口函数

# -*- coding: utf-8 -*- # @Time : 2023/10/13 15:08 # @Author : RaSo # @File : interface_keypoints_for_standingDetect.py # Begin to show your code! from ultralytics import YOLO import math # 本接口输入和yolov8一样，输出为目标框和站姿检测结果。 # 实现方式，首先，通过yolov8-pose模型，预测出目标和关键点。输出为目标类别和box，以及该目标17个关键点（x,y,score） # 然后，左右肩，髋，膝，脚踝，肢体关系，确定站姿的目标，输出人的目标框和站姿检测结果output= 图像，{person：standYes/No} def calculate_distance(x1, y1, x2, y2): return math.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2) # B相对于点A和点C的夹角 def calculate_angle(A, B, C): AB = calculate_distance(A[0], A[1], B[0], B[1]) BC = calculate_distance(B[0], B[1], C[0], C[1]) AC = calculate_distance(A[0], A[1], C[0], C[1]) cos_angle = (AB ** 2 + BC ** 2 - AC ** 2) / (2 * AB * BC) # 余弦 angle = math.degrees(math.acos(cos_angle)) return angle, cos_angle # 返回置信度在0.5以上的关键点个数，用来判断计算结果可信与否。 def count_numbers_greater_than_0_5(input_list): count = 0 for score in input_list: if score > 0.60: count += 1 return count # 测试目标lineheights def testStandYesNo_withKeypoint(mode_pt=None, source=None): # device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") kpt_shape: [17, 3] # number of keypoints, number of dims (2 for x,y or 3 for x,y,visible) flip_idx: [0, 2, 1, 4, 3, 6, 5, 8, 7, 10, 9, 12, 11, 14, 13, 16, 15] # 输出关键点顺序 # 关键点类别索引 kpClasses = ['nose', 'left_eye', 'right_eye', 'left_ear', 'right_ear', 'left_shoulder', 'right_shoulder', 'eft_elbow', 'right_elbow', 'left_wrist', 'right_wrist', 'left_hip', 'right_hip', 'left_knee', 'right_knee', 'left_ankle', 'right_ankle'] model = YOLO(model=mode_pt) # 加载模型 # model.to(device) result = model.predict(source=source, save=True) standYesNoResult = dict() # 保存输出结果 # 遍历关键点模型输出结果 for imgres in result: print('\n', imgres.path) everyImg = [] for keyP, personBox in zip(imgres.keypoints, imgres.boxes): kpList = keyP.data.tolist() personBoxList = personBox.data.tolist() # 关键点，左右肩部，左右臀部，左右膝盖，左右踝关节 leftShoulder = kpList[0][5] # 肩 rightShoulder = kpList[0][6] leftHip = kpList[0][11] # 髋 rightHip = kpList[0][12] leftKnee = kpList[0][13] # 膝盖 rightKnee = kpList[0][14] leftAnkle = kpList[0][15] # 脚踝 rightAnkle = kpList[0][16] # 关键点肢体夹角（髋处）与余弦，和存在与否置信度得分影响因子 angle1, cos1 = calculate_angle(leftShoulder, leftHip, leftKnee) angle2, cos2 = calculate_angle(rightShoulder, rightHip, rightKnee) real1 = count_numbers_greater_than_0_5([leftShoulder[-1], leftHip[-1], leftKnee[-1]]) real2 = count_numbers_greater_than_0_5([rightShoulder[-1], rightHip[-1], rightKnee[-1]]) # 关键点肢体夹角（膝盖处）与余弦，和存在与否置信度得分影响因子 angle3, cos3 = calculate_angle(leftHip, leftKnee, leftAnkle) angle4, cos4 = calculate_angle(rightHip, rightKnee, rightAnkle) real3 = count_numbers_greater_than_0_5([leftHip[-1], leftKnee[-1], leftAnkle[-1]]) real4 = count_numbers_greater_than_0_5([rightHip[-1], rightKnee[-1], rightAnkle[-1]]) print(cos1, cos2, cos3, cos4, real1, real2, real3, real4, 'person=', personBoxList) # 姿势判断，是否站立 # 1.如果两个膝关节都是弯的，就不可能站立姿态。 if real3 == 3 and real4 == 3 and cos3 > -0.2 and cos4 > -0.2: everyImg.append( {'boxx': personBoxList, 'standingYN': 0, 'Description': "No standing"}) elif ((real1 == 3 and cos1 < -0.88) and (real2 == 3 and cos2 < -0.88)) or \ (((real1 == 3 and cos1 < -0.88) or (real2 == 3 and cos2 < -0.88)) and ( 4 <= real3 + real4 <= 5) and cos3 < -0.707 and cos4 < -0.707): # 髋关节夹角的六个点都显现，置信度高。则高要求。 print('Standing-person=', personBoxList, ',there is a standing person', 'cos=', min(cos1, cos2)) everyImg.append( {'boxx': personBoxList, 'standingYN': 1, 'Description': "Is standing"}) elif (real1 == 3 and real2 <= 2 and cos1 < -0.88) or ( real1 <= 2 and real2 == 3 and cos2 < -0.88): # 组成髋关节的三个点至少应该有一组是完整的，另一组不完整时有此表达式。 print('Standing-person=', personBoxList, ',there is a standing person', 'cos=', min(cos1, cos2)) everyImg.append( {'boxx': personBoxList, 'standingYN': 1, 'Description': "Is standing"}) else: everyImg.append( {'boxx': personBoxList, 'standingYN': 0, 'Description': "No standing"}) standYesNoResult[imgres.path] = everyImg print("over") return standYesNoResult if __name__ == '__main__': mode_pt = r'D:/TanHaiyan/Models/yolov8/pose/yolov8l-pose.pt' source = r"D:/TanHaiyan/Datasets/keypoints_pose2023/testKeypoint/" res = testStandYesNo_withKeypoint(mode_pt, source) print(res)