Human Activity Detection Using Phone and Watch Sensors

Dataset

WISDM Smartphone and Smartwatch Activity and Biometrics Dataset

Dataset Summary

Metric	Value
Number of subjects	51
Number of activities	18
Minutes collected per activity	3
Sensorspolling rate	20Hz
Smartphone used	Google Nexus 5/5X or Samsung Galaxy S5
Smartwatch used	LG G Watch
Number of raw measurements	15,630,426

18 activity are recorded and coded as follow

Activity	Code
Walking	A
Jogging	B
Stairs	C
Sitting	D
Standing	E
Typing	F
Brushing Teeth	G
Eating Soup	H
Eating Chips	I
Eating Pasta	J
Drinking from Cup	K
Eating Sandwich	L
Kicking (Soccer Ball)	M
Playing Catch w/Tennis Ball	O
Dribbling (Basketball)	P
Writing	Q
Clapping	R
Folding Clothes	S

Prerequisites :

python >= 3.8

pip install -r requirements.txt

Data Preprocessing

code/001_data_collection_preprocessing_phone_data.ipynb
code/001_data_collection_preprocessing_watch_data.ipynb

EDA

code/002_eda_phone_data.ipynb
code/002_eda_watch_data.ipynb

Feature extraction

Performs signal processing and generates processed dataset in dataset/ folder

code/003_feature_extraction.ipynb

Training

Note: It is prerequisite to execute above feature extraction since Logistic regression, Random forest and KNN works on feature extracted dataset.

1. Logistic regression

code/004_1_train_logisticregression_phone.ipynb
code/004_1_train_logisticregression_watch.ipynb

2. KNN

code/004_3_train_knn_phone.ipynb
code/004_3_train_knn_watch.ipynb

3. Random forest

code/004_2_train_randomforest_phone.ipynb
code/004_2_train_randomforest_watch.ipynb
code/004_2_train_randomforest.ipynb

4. LSTM

code/004_4_train_lstm_phone.ipynb
code/004_4_train_lstm_watch.ipynb
code/004_4_train_lstm.ipynb

5. GRU

code/004_5_train_gru_phone.ipynb
code/004_5_train_gru_watch.ipynb
code/004_5_train_gru.ipynb

Results

Single Sensor

1. Logistic regression

Sensors	Accuracy(%)
Phone Accelerometer	39.57
Phone Gyroscope	40.56
Watch Accelerometer	73.06
Watch Gyroscope	65.08

2. KNN

Sensors	Accuracy(%)
Phone Accelerometer	59.29
Phone Gyroscope	44.20
Watch Accelerometer	62.64
Watch Gyroscope	57.03

3. Random forest

Sensors	Accuracy(%)
Phone Accelerometer	72.99
Phone Gyroscope	63.81
Watch Accelerometer	85.60
Watch Gyroscope	76.60

4. LSTM

Sensors	Accuracy(%)
Phone Accelerometer	25.75
Phone Gyroscope	49.37
Watch Accelerometer	74.97
Watch Gyroscope	76.57

5. GRU

Sensors	Accuracy(%)
Phone Accelerometer	30.10
Phone Gyroscope	53.03
Watch Accelerometer	78.38
Watch Gyroscope	80.80

Multi Sensors

1. Random forest

Sensors	Accuracy(%)
Phone Accel & Gyro	81.04
Watch Accel & Gyro	90.26
All 4 Sensors	93.17
All 4 Sensors (15 activities)¹	96.12

2. LSTM

Sensors	Accuracy(%)
Phone Accel & Gyro	52.79
Watch Accel & Gyro	85.19
All 4 Sensors	85.11
All 4 Sensors (15 activities)¹	89.32

3. GRU

Sensors	Accuracy(%)
Phone Accel & Gyro	58.18
Watch Accel & Gyro	88.14
All 4 Sensors	88.06
All 4 Sensors (15 activities)¹	92.27

¹ dropping "eating chips", "eating pasta", "eating sandwich" activities.

Conclusion

This project focused on analysing the WISDM dataset to classify human activities using a range of machine learning and deep learning models. Among the tested approaches, the Random Forest classifier stood out as the top performer, delivering the highest accuracy and computational efficiency. Its robust handling of feature importance and generalization across the dataset made it a reliable choice for this task.

Deep learning models, particularly LSTM and GRU, also showed strong potential. These models worked well in capturing temporal patterns and dependencies inherent in sequential data, showing their suitability for activity recognition tasks. However, further optimization through hyperparameter tuning and architectural adjustments could enhance their performance and fully leverage their strengths.

Future directions for this work could include integrating ensemble techniques to combine the strengths of different models, developing hybrid frameworks that utilize both traditional machine learning and deep learning, and employing more advanced preprocessing steps to improve data quality. These efforts could drive better performance, greater adaptability and deeper insights into human activity classification.

Code Contributors

        1) Murthy L
        2) Moushumi Mahato
        3) Shashankgoud Patil