fall detection nicholas chan (ee) abhishek chandrasekhar (ee) hahnming lee (ee) akshay patel (cmpe)...

34
Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Upload: aleesha-fox

Post on 17-Dec-2015

218 views

Category:

Documents


3 download

TRANSCRIPT

Page 1: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Fall Detection

Nicholas Chan (EE)Abhishek Chandrasekhar (EE)

Hahnming Lee (EE)Akshay Patel (CmpE)

1

Page 2: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Elderly Fall Statistics

16,000 elderly Americans die from falling each year (CDC, 2005)

300,000 elderly Americans have hip fractures each year

90% of hip fractures result from falls

24% of elderly Americans who suffer hip fractures die within one year

40% of elderly women with hip fractures never walk unassisted again (National Osteoporosis Foundation)

2

Page 3: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Proposed Solution

Two camera system executing custom algorithm:

1. Detect person in room

2. Perform statistical analysis of person’s motion

3. Determine if a fall has occurred

4. Send an alarm for help

Projected cost of $500 per room

3

Page 4: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Target Market

4

Smart Hospital Rooms Nursing Homes & Clinics

Our solution offers to reduce injuries arising from falls and to improve safety records at nursing homes and hospitals.

Page 5: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Alternative Solutions

Pressure-sensitive mats by the bed

Camera detection with optical flow algorithm

RFID Solutions

Accelerometers (e.g., iLife ™)

5

Page 6: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Alternative Solution Problems

Pressure sensitive mats have unavoidable edges that can cause falls

Optical flow analysis prone to errors arising from shadow artifacts

Potential EMI interference from RFID readers; RFID readers also very expensive (over $1000)

Accelerometer results in many false positives (e.g. a person sitting down quickly)

6

Page 7: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Technical Specifications

Two webcams (Microsoft VX 6000) Resolution of 160x120 pixels Video recorded at 15 frames per second

Personal Computer to run algorithm: Intel Pentium Dual Core 2.5GHz Processor 3GB RAM Standard Keyboard and Mouse

7

Page 8: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Camera Positioning

Privacy is a major concern

Gaining maximal coverage from camera position is also critical

A balance between these two must be achieved

8

Page 9: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Camera PositioningMaximal Coverage

Head-level Camera

High-level camera

Coverage Area

9

Page 10: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Camera PositioningMaximal Privacy

High-level camera

Knee-level camera Coverage Area

10

Page 11: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Algorithm Overview

1. Identify the region of an image occupied by the person

2. Ascertain the velocity of the person’s motion

3. Fit an ellipse to the person

4. Analyze the changes in the ellipses’ properties

5. Determine if a fall has occurred

11

Page 12: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Foreground Segmentation

The background of every frame is subtracted

Statistical Gaussian model is generated for each pixel

HSV color space is used to minimize shadow effect

Pixels are labeled as either foreground or background based on a preset threshold

A binary foreground image is thus generated

12

Page 13: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Foreground Segmentation

13

Foreground Segmentation

Foreground Segmentation

Page 14: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Foreground Segmentation

14

Foreground Segmentation

Page 15: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Largest Blob Detection

Additional filtering is performed on the foreground-segmented image

The largest continuous cluster of pixels is detected and then isolated from the smaller clusters of noise

15

Page 16: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Largest Blob Detection

Blob Detection

Blob Detection

16

Page 17: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Motion History Imaging

17

Filtered foreground-segmented image data used to form Motion History Image (MHI)

MHI used to quantify the velocity of the person’s motion

0 (zero velocity) ≤ Cmotion ≤ 1 (extreme velocity)

Page 18: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Motion History Imaging

Swiftly Walking( Medium Cmotion )

Turning Around( Low Cmotion )

Falling( High Cmotion )

Cmotion numbergray pixels

numbergray number white pixels18

Page 19: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Elliptical Approximation

19

Frame 1

Normal Walking

Frame 150

Mid-FallChange in

Ellipse Angle

Page 20: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Frame 120

Normal Walking

Frame 150

Mid-FallChange in

Eccentricity20

Elliptical Approximation

Page 21: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

0 50 100 150 200 250 300 350 400 4500

0.5

1

1.5

2

Frame

Elli

pse

Ang

le (r

ad)

Angle

0 50 100 150 200 250 300 350 400 4501

2

3

4

Frame

Elli

pse

Ecc

entri

city

Eccentricity

High Frequency Noise

Possible Fall

Elliptical Approximation

Page 22: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Statistical Analysis

Falls result in: 1) high-velocity motion (high Cmotion values) and2) large statistical variance in elliptical orientation/eccentricity

Numerically, we define a fall is defined by:Cmotion > 0.65 and σθ > 0.60

These thresholds may vary slightly with camera position

22

Page 23: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Statistical Analysis

23

Cmotion > 0.65

σθ > 0.60

Page 24: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Call for Assistance

Computer connected to Ethernet network

When fall happens a picture is taken

A fuzzy picture is stored to a local server

An updating intranet page is displayed at the nurse station

The page incorporates archiving features

Nurse analyzes picture and determines if a response is necessary

24

Page 25: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Call for Assistance UI

25

Page refreshes every 5 seconds to check for screenshot on the server

Page 26: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Call for Assistance UI

26

When a fall occurs a flashing red message along with a screenshot is displayed

Page 27: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Archiving Falls

27

The shot can be archived with a date stamp onto the local server

The detected fall log shows a queue of falls that happened

On archiving and reloading the system shows normal status again

Page 28: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Results

Results are based on evaluation of 30 falls and 20 non-falls

28

Category % Success % Failure

Falls 83.33 % 16.66 %

Non-Falls 75 % 25 %

Page 29: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Problems and Solutions

Hardware and Software Problems: MATLAB requires substantial memory to execute programs Algorithm has difficulty accounting for auto-light adjustments

by the webcam

Solutions Proposed: Port existing algorithm to C++ in order to run it more

efficiently; using C++ also removes the licensing hassles required with MATLAB

Light intensity can be normalized with histogram equalization techniques; alternatively use a webcam without light adjustment

29

Page 30: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Real-Time Analysis

Existing Problems: MATLAB is incapable of running threaded applications Analysis and recording of video simultaneously is

almost impossible as a result

Solution: Use C++; Supports threading and memory

management Real time analysis is available via OpenCV library Many MATLAB functions are implemented in the library

30

Page 31: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Privacy Concerns

Use of cameras brings in a major privacy concern

Different configurations are necessary for concealment

Terms & Conditions have to be included in hospital paperwork

The picture taken of the patient upon a fall is blurred

An option of not having the system on should be implemented if requested by the patient

31

Page 32: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Cost Analysis

Assuming a rate of $28/hr, Engineer salaries would amount to $44,800 for 4 engineers during a 10 week development phase

Equipment Cost: $60 for two cameras $270 for a modern Dell Inspiron 530

$170 Installation and Software Costs

Total Cost per Room = $500

32

Page 33: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Future Improvements

Enable support for multiple people

Improve speed of algorithm

Reduce false positives by making a self-learning system

Make the program standalone for easy deployment

Enable mainframe support for hospital with servers

33

Page 34: Fall Detection Nicholas Chan (EE) Abhishek Chandrasekhar (EE) Hahnming Lee (EE) Akshay Patel (CmpE) 1

Questions?

16,000 Americans die from falling each year

300,000 elderly Americans have hip fractures each year

24% elderly Americans who suffer hip fractures die within one year

34

Category % Success % Failure

Falls 83.33 % 16.66 %

Non-Falls 75 % 25 %