HAPT-Recognition

Human Activities and Postural Transitions’ Recognition using Smartphone Data

PROBLEM STATEMENT DESCRIPTION

Human activities are monitored with the help of Smartphone sensors(Acclerometer and Gyroscope). The statement is to classify the human activities into one of 12 classes based on these sensor readings.

DATASET

Smartphone-Based Recognition of Human Activities and Postural Transitions Data Set

The smartphone sensor data are transformed into two categories:-

• Time Domain Features - Acclearation(x,y,x), min, median, entropy, etc.

• Frequency Domain Features - DFT of time domain features(accleration, jerk magnitude, gyroscope magnitude, etc).

Data Visualization

Data is visualized using 2-D PCA and TSNE embeddings. TSNE visualization shows that the different classes are well seperable.

EXPERIMENTS

Several classification techniques are implemented across different parameter variation. A detailed study of all the experiment as mentioned below are presented:

• Neural Network(Single and Multilayer perceptron)
• SVM(Linear and Gaussian Kernel)
• Boosting(with different loss functions)

Single Layer Neural Network

Training Accuracy(%)	Validation Accuracy(%)	Test Accuracy(%)
97.55	96.2	92.13

Multilayer Neural Network

Number of hidden units	Training Accuracy(%)	Validation Accuracy(%)	Test Accuracy(%)
128	98.28	97.17	93.17
256	99.03	97.04	93.39
512	99.51	97.94	93.48

L2- SVM

Kernel	Parameters	Training Accuracy(%)	Validation Accuracy(%)	Test Accuracy(%)
Linear	C = 1	99.53	96.98	95.19
Gaussian	C = 5000, gamma = 1e-5	98.83	96.6	94.4

Boosting

Loss Function	Weak learners	Number of weak learner	Training Accuracy(%)	Validation Accuracy(%)	Test Accuracy(%)
Exponential	Decision Stumps	339	99.97	95.6	91.68
Cross Entropy	Decision Stumps	303	99.41	94.21	91.4