yueyrhansdfeddjrjrje9ojdmneojkrjfjei.pptx

Spiritual pdf file attached

1. Digital Access for Community Empowerment - 2 Elderly Fall Detection System A Muhammad Thanzeem Archana Thomas Athira C Goutham Krishnan P Mevin Mathew Rakesh Ramachandran

2. Table of contents 01 04 02 05 03 06 Inspiration Block Diagram Data Acquisition Data Training and testing Hardware Implementation Future Scope ‹#›

3. Our Inspiration Fall among elderly has always been a concerning issue, as it could cause permanent impairments or even death.This project attempts to build a real-time fall detection system with an app interface that will help the caretakers/dependants with a warning message. ‹#›

4. MPU6050 The MPU6050 ‹#›

5. Data Acquisition 1 Data Acquisition using Arduino Uno, MPU6050. 3 Implement the model to ESP32 and STM32l433rc-p Nucleo board 4 Setup the PCB and the design the enclosure for the prototype Upload the acquired data and mobifall data and train the model using tensorflow. 2 Data Training and Testing Model Implementation Hardware Setup ‹#›

6. Data Acquisition A set of data is taken by sample falls by each of the team members. These include falls from standing and walking positions to the ground. Team members have fell on side, back, knees and front with accelerometer sensor around their waist. Live from team members Mobifall Dataset The major training data is taken from the mobifall dataset which includes falls from different positions and orientations like fall from stairs climbing up, fall from a sitting chair, fall from walking etc. 01 ‹#›

7. Brief Theory Model Theory Equations Application in Fall detection k-Nearest Neighbors (k-NN) Non-parametric method for classification and regression. Classifies data points based on the majority class among k nearest neighbors. Classifies accelerometer readings as "fall" or "no fall" based on proximity to labeled examples. Neural Networks (NN) Composed of interconnected layers of neurons. Each neuron applies a weighted sum of its inputs and passes it through an activation function. Learns complex patterns in accelerometer data, distinguishing between fall and non-fall events. ‹#›

8. Theory Model Theory Equations Application in Fall detection MFCC Represents the power spectrum of a signal in a compact form. Used primarily in audio and speech processing. Captures short-term spectral characteristics. MFCCs can be used as features in machine learning models to classify accelerometer data, as they capture significant frequency components relevant to detecting falls. Gradient Boosting Ensemble learning technique that builds models sequentially, each new model aiming to correct the residual errors of the previous ones. Improves fall detection accuracy by iteratively correcting prediction errors and focusing on difficult-to-classify instances. ‹#›

9. Theory Model Theory Equations Application in Fall detection Support Vector Machine (SVM) Finds the optimal hyperplane separating different classes in the feature space. Aims to maximize the margin between the classes. Classifies accelerometer data into fall and no-fall categories by finding an optimal separating hyperplane, even in complex feature spaces. Random Forest Ensemble learning method that combines multiple decision trees to improve classification accuracy and robustness. provides accurate fall detection by aggregating predictions from multiple decision trees and assessing feature importance. Hidden Markov Model (HMM) Statistical models representing systems with hidden states, useful for seque n tial data analysis. HMMs can model sequences of accelerometer data to identify patterns indicative of falls based on transitions between hidden states. ‹#›

10. Data Acquisition 01 ‹#›

11. FALL SET 1 FALL SET 2 FALL SET 3 REAL TIME FALL SAMPLES ‹#›

12. Forward Fall Walking Sitting in chair Variability in Falls Human Variability Feature Extraction Continuous Learning FALL SAMPLES FROM DATASET ‹#›

13. FALL DATA ‹#›

14. NO FALL ‹#›

15. Model Predictions with neural network ‹#›

16. Model Parameters ‹#›

17. Real Time Prediction and Response Time - neural network ‹#›

18. Feature extraction with MFCC SVM KNN Gradient Boosting Random Forest ‹#›

19. MFCC-Feature extraction - 32 cepstral features SVM KNN Note : Accuracy is seen to decrease when number of cepstral features are increased. So we did hypertuning to find the parameters that gave the best accuracy. ‹#›

20. SVM Best parameters Sl.No Model param Values 1 SVM_C 0.1 2 SVM_gamma 0.001 3 SVM_kernel poly 4 SVM_degree 2 Sl.No MFCC param Values 1 No of cepstral features 26 2 Window length .512 3 Window step .1 ‹#› Accuracy = 69.44%

21. Test accuracy vs window length Test accuracy vs number of cepstral features. ‹#›

22. Random Forrest Best parameters Sl.No Model param Values 1 rf_max_depth none 2 rf_min_samples_leaf 1 3 rf_min_samples_split 2 4 rf_n_estimators 100 Sl.No MFCC param Values 1 No of cepstral features 20 2 Window length 1 3 Window step .1 ‹#› Accuracy = 88.88%

23. Random Forrest Test accuracy vs window size Test accuracy vs window step Test accuracy vs no of mfcc features ‹#›

24. KNN Best parameters Sl.No Model param Values 1 n neighbours 3 2 weights distance 3 algorithm auto Sl.No MFCC param Values 1 No of cepstral features 13 2 Window length .025 3 Window step .512 ‹#› Accuracy = 66.66%

25. Test accuracy vs No of MFCC Features Test accuracy vs window size Test accuracy vs window step 1) 2 ) 3 ) ‹#›

26. Gradient Boosting Best parameters Sl.No Model param Values 1 n estimators 50 2 learning rate 0.2 3 max depth 3 Sl.No MFCC param Values 1 No of cepstral features 13 2 Window length .1 28 3 Window step .01 ‹#› Accuracy = 94.4%

27. HMM Best parameters Sl.No Model param Values 1 n variance 2 2 covariance type diag Sl.No MFCC param Values 1 No of cepstral features 13 2 Window length .025 3 Window step .01 ‹#› Accuracy = 63.88%

28. SVM Best parameters - 10 fold cross validation Sl.No Model param Values 1 SVM_C 10 2 SVM_gamma 0.0001 3 SVM_kernel rbf 4 SVM_degree 2 Sl.No MFCC param Values 1 No of cepstral features 26 2 Window length 1 3 Window step .1 ‹#›

29. Random Forrest - 10 fold CV Sl.No Model param Values 1 rf_max_depth none 2 rf_min_samples_leaf 1 3 rf_min_samples_split 2 4 rf_n_estimators 100 Sl.No MFCC param Values 1 No of cepstral features 20 2 Window length 1 3 Window step .1 ‹#›

30. KNN Best parameters - 10 fold CV Sl.No Model param Values 1 n neighnours 3 2 weights uniform 3 algorithm auto Sl.No MFCC param Values 1 No of cepstral features 26 2 Window length .025 3 Window step .512 ‹#›

31. Gradient Boosting Best parameters Sl.No Model param Values 1 n estimators 1 50 2 learning rate 0.1 3 max depth 3 Sl.No MFCC param Values 1 No of cepstral features 13 2 Window length .025 3 Window step .01 ‹#›

32. HMM Best parameters - 10 fold CV Sl.No Model param Values 1 n variance 2 2 covariance type diag Sl.No MFCC param Values 1 No of cepstral features 13 2 Window length .025 3 Window step .01 ‹#›

33. Comparison Best test accuracy found while using SVM. ‹#›

34. Implementation in STM Board ‹#› SVM Random Forrest Gradient Boosting MACC 3146000 896 1200 Flash (KB) 1270 31.328 32.228 RAM (KB) 12.224 10.40 7.240 Elapsed Time (s) 1.835 1.107 2.978

35. Real Time Prediction using random forest ‹#›

36. MPU6050 STM32 Microcontroller App Interface MPU6050 STM - Nucleo LCD Screen GSM Module Final Product PROTOTYPE ‹#›

37. MODEL IMPLEMENTATION 03 WITH STM - NUCLEO BOARD ‹#›

38. MODEL IMPLEMENTATION 03 WITH ESP-32 BOARD ‹#›

39. PCB DESIGN WITH ESP32 04 ‹#›

40. HARDWARE ENCLOSURE 04 ‹#›

41. PCB DESIGN WITH STM32 MiCROCONTROLLER 04 ‹#›

42. FRONT-END INTERFACE 05 ‹#›

43. FRONT-END INTERFACE 05 ‹#›

44. FRONT-END INTERFACE 05 ‹#›

45. CONCLUSION ‹#› Significance of Fall Detection : Crucial for ensuring safety and well-being of the elderly. Timely detection reduces serious injuries and healthcare costs. Technological Approaches : Wearable sensors Sensor Fusion and Advanced Algorithms : Combining sensor data with advanced algorithms improves accuracy and reliability. Challenges and Future Directions : Addressing false alarms, and real-time processing remains essential. Focus on enhancing algorithm performance, reducing power consumption, and improving user-friendly designs.

46. FUTURE SCOPE 05 Seemingless integration of STM32l433rc-p Nucleo, MPU6050 and GSM/WiFi module for immediate transfer of fall alarm to the caretakers. Optimized data training and testing techniques where an alarm is sent if the device senses when a person is about to fall. ‹#›

47. THANK YOU ‹#›

48. Component Specification MPU6050 03 ‹#›

49. Component Specification STM32l433rc-p Nucleo 03 ‹#›

50. Memory utilization and power consumption Note : Power consumption is noted only for gradient boosting ‹#›

51. Best MFCC Parameters ‹#›

52. Hidden Markov model Note : Need to work on accuracy level of using this model ‹#›

53. BEST PARAMETERS ‹#›

54. Comparison ‹#›

55. BILL OF MATERIAL ‹#›