tools.py

import math
import os
import urllib
import pandas as pd
from PIL import Image
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score
from poseEstimation import image_to_pose
from tqdm import tqdm
from main import fall_detection

FALL = 1
NOT_FALL = 0


def get_data_from_images(folder_path):
    x = []
    y = []
    for folder in os.listdir(folder_path):
        for file in tqdm(os.listdir(folder_path + "/" + folder)):
            pose = image_to_pose(folder_path + "/" + folder + "/" + file)
            if len(pose) >= 1:
                x.append(pose)
                y.append(FALL if folder == "fall" else NOT_FALL)
    return pd.DataFrame({"x": x, "y": y})


def store_data(df, folder_path):
    df.to_csv(folder_path + "/data.csv", index=False)
    df.to_pickle(folder_path + "/data.pkl")


def load_data(folder_path):
    return pd.read_pickle(folder_path + "/data.pkl")


def check_model_on_data(from_images=False):
    if from_images:
        data = get_data_from_images("fall_dataset/old")
        store_data(data, "fall_dataset/data")
    else:
        data = load_data("fall_dataset/data")
    dataset = data.values
    x = dataset[:, 0]
    y = dataset[:, 1]
    y_predict = []
    for i in range(len(x)):
        y_predict.append(int(fall_detection(x[i])[0]))

    y_predict = list(y_predict)
    y = list(y)

    accuracy = accuracy_score(y, y_predict)
    print("Accuracy: %f" % accuracy)
    precision = precision_score(y, y_predict)
    print("Precision: %f" % precision)
    recall = recall_score(y, y_predict)
    print("Recall: %f" % recall)
    f1 = f1_score(y, y_predict)
    print("F1 score: %f" % f1)


def crop_humans_from_image(image_path, save_path="fall_dataset/images/cropped_images"):
    img = Image.open(image_path)
    img_name = "-".join(image_path.split("/")[-1].split(".")[:-1])
    img_type = image_path.split("/")[-1].split(".")[-1]
    w, h = img.size
    new_w, new_h = 960, int((960 / w) * h)

    pose = image_to_pose(image_path)
    for idx in range(len(pose)):
        xmin, ymin = (pose[idx, 2] - pose[idx, 4] / 2), (
            pose[idx, 3] - pose[idx, 5] / 2
        )
        xmax, ymax = (pose[idx, 2] + pose[idx, 4] / 2), (
            pose[idx, 3] + pose[idx, 5] / 2
        )
        xmin, ymin = int(xmin * w / new_w), int(ymin * h / new_h)
        xmax, ymax = int(xmax * w / new_w), int(ymax * h / new_h)
        xmin, ymin = xmin - 5, ymin - 5
        xmax, ymax = xmax + 5, ymax + 5
        crop = img.crop((xmin, ymin, xmax, ymax))
        crop.save(save_path + "/" + img_name + "_cropped_" + str(idx) + "." + img_type)


# Clean the images. The function will crop and save the persons in the image into new images.
def clean_images(folder_path, save_path):
    for file in tqdm(os.listdir(folder_path)):
        crop_humans_from_image(folder_path + "/" + file)


# Predict if the person in the image is falling or not. The image must be a web-url.
def predict(image_url):
    urllib.request.urlretrieve(image_url, "image.png")
    poses = image_to_pose("image.png")
    for pose in poses:
        if fall_detection(pose):
            return True
    return False


# predict('https://st2.depositphotos.com/1000393/9807/i/950/depositphotos_98078022-stock-photo-man-falling-down.jpg')
check_model_on_data(False)
# clean_images('fall_dataset/images/fall', 'fall_dataset/images/cropped_images')
# clean_images('fall_dataset/images/not-fall', 'fall_dataset/images/cropped_images')