1 of 9

mmWave-based

Activity Recognition

Soo Min Kwon

Shreya Patel

Christine Mathews

Wesam Saleh

Allen Zhang

1

Advisor: Dr. Yingying Chen

2 of 9

PREVIOUSLY...

  • Implemented an SVM classifier with training data that had 5 different features: mean, median, standard deviation, skewness, and kurtosis
  • Looked into TensorFlow in attempt to set up a CNN environment for improvement
  • Researched related papers

2

3 of 9

Testing SVM Classifier

3

Training Dataset for Walking Towards (Shreya)

Testing Dataset for Walking Towards (Christine)

Correct

Prediction

Results

(Label 0 for Walking Towards)

4 of 9

Challenges with SVM Classifier

​

  • Can only detect activities that we experimented and trained the classifier with
    • General classification problem
    • Needs lots of experiments and datasets to train classifier with
    • Can possibly make mistakes with other people that the classifier has not been trained with yet

4

5 of 9

New Reference Papers & Related Works

  1. Can WiFi Estimate Person Pose?
    1. This paper, in short, uses WiFi signals and CSI data to further analyze and make “skeletal” results for human poses.
  2. 2D/3D Pose Estimation and Action Recognition using Multitask Deep Learning
    • This paper tries to use the pose estimation results to classify human activity recognition.

5

6 of 9

Can WiFi Signals Estimate Person Pose?

6

7 of 9

2D/3D Pose Estimation and Action Recognition using Multitask Deep Learning

7

  • Get 2D/3D estimated poses from RGB images
  • Get other visual features from Multitask CNN from the RGB images
  • Use aggregator/classifier to predict human activities

8 of 9

THIS WEEK & NEXT WEEK

  • Finish reading related papers on pose estimation
  • Learn on setting up a convolutional neural network environment
  • Try to get an estimated human pose based on existing work

8

9 of 9

Thank You!

​

Q&A

9