Skip to content

Latest commit

 

History

History
35 lines (23 loc) · 1.4 KB

README.md

File metadata and controls

35 lines (23 loc) · 1.4 KB
Raw

Reference

Papers and Strategies

  1. A Flexible Fall Detection Framework Based on Object Detection and Motion Analysis

    First detects human objects in the frame using YOLOv5, then tracks the object by using YOLOv5 every frame.

  2. Improved Pedestrian Fall Detection Model Based on YOLOv5

    Doesn't use motion analysis, but uses YOLOv5 and a CNN to detect pose.

  3. Fall Detection with Pose Estimation

  4. also pose estimation calculating hip position

Proposed Strategy

  1. Detect people in the frame
    • YOLOV5
    • Tensorflow pose detection (movenet)
  2. For each estimated pose:
    • calculate the angle of the head and the hip
      • calclulate the rotation speed
      • calculate the angular acceleration
    • calculate the position of the hip
      • calculate the difference between y velocity and acceleration
    • calculate the width and height ratio of the bounding box
  3. If the angular acceleration of the head and the hip is greater than a threshold, and the difference between y velocity and acceleration is greater than a threshold, and the width > height, then the person is falling.

Dataset

  1. Multiple cameras fall dataset