根据图片提取人体关节角数据

收集一定数量的人体关节角度，通过相应的姿态分类算法，用于判断分类器准确与否通过pandas库读取数据集图像，再调用movenet thunder模型估计得出人体骨骼关键点在图片中的位置（关键点坐标）和置信度，选用坐标置信度最高的17个坐标作为关键点坐标，：鼻子，左/右眼，左/右耳朵，左/右肩膀，左/右手肘，左/右手腕，左/右臀部，左/右膝盖，左/右脚踝。由于人高矮胖瘦不同，需要采取一些视不变特征作为数据输入，这里根据关键点, 计算十个关节角，左右头部关节角：鼻子坐标A，耳朵坐标B，肩膀坐标C；左右肩膀关节角：肩膀坐标A，手肘坐标B，臀部坐标C；左右背部关节角：臀部角度A，肩膀角度B，膝盖角度C；左右手肘关节角：将手肘坐标A，手腕坐标B，肩部坐标C；左右膝关节角：膝关节坐标A，臀部坐标B，脚踝坐标C输入，计算各边长，通过三角余弦公式cos(A) = (b^2 + c^2 - a^2) / (2bc)计算人体关节角度的余弦值作为视不变特征，将add_train.csv输入svm，knn等分类器训练，在将add_test.csv输入训练好的svm，knn等分类器，将分类结果与class no对比准确率

Dataset Construction and Evaluation Process: First, read the dataset images using the pandas library. Next, utilize the MoveNet Thunder model to estimate the coordinates and confidence scores of human skeletal keypoints in the images, and select the 17 keypoints with the highest confidence scores, including: nose, left/right eye, left/right ear, left/right shoulder, left/right elbow, left/right wrist, left/right hip, left/right knee, and left/right ankle. To mitigate the impact of variations in human physique (height, weight, etc.), scale-invariant features are adopted as model inputs. Based on the selected keypoints, ten joint angles are calculated: left/right head joint angles (using coordinates of the nose, ear, and shoulder), left/right shoulder joint angles (using coordinates of the shoulder, elbow, and hip), left/right back joint angles (using coordinates of the hip, shoulder, and knee), left/right elbow joint angles (using coordinates of the elbow, wrist, and shoulder), and left/right knee joint angles (using coordinates of the knee, hip, and ankle). The side lengths of the corresponding triangles are computed, and the cosine values of the human joint angles are derived via the trigonometric cosine formula: cos(A) = (b² + c² - a²)/(2bc), which serve as the scale-invariant features. Subsequently, classifiers such as SVM and KNN are trained using the add_train.csv dataset. Finally, the add_test.csv dataset is input into the trained classifiers, and the classification accuracy is calculated by comparing the classification results with the class labels (class no) to assess the performance of the classifiers.