RadHAR: Human Activity Recognition from Point Clouds Generated through a Millimeter-wave Radar
This repo contains the implementation for the paper: RadHAR: Human Activity Recognition from Point Clouds Generated through a Millimeter-wave Radar presented at the 3rd ACM Workshop on Millimeter-Wave Networks and Sensing Systems (mmNets) 2019 (co-located with MobiCom).
Akash Deep Singh, Sandeep Singh Sandha, Luis Garcia, Mani Srivastava
Point cloud dataset collected using mmWave Radar. The data is collected using the ROS package for TI mmWave radar developed by radar-lab. Raw dataset is available in the repo. The data is partitioned into the folders train and test. Each train and test folder further contains the folders which have respective activity classes.
header:
seq: 6264
stamp:
secs: 1538888235
nsecs: 712113897
frame_id: "ti_mmwave" # Frame ID used for multi-sensor scenarios
point_id: 17 # Point ID of the detecting frame (Every frame starts with 0)
x: 8.650390625 # Point x coordinates in m (front from antenna)
y: 6.92578125 # Point y coordinates in m (left/right from antenna, right positive)
z: 0.0 # Point z coordinates in m (up/down from antenna, up positive)
range: 11.067276001 # Radar measured range in m
velocity: 0.0 # Radar measured range rate in m/s
doppler_bin: 8 # Doppler bin location of the point (total bins = num of chirps)
bearing: 38.6818885803 # Radar measured angle in degrees (right positive)
intensity: 13.6172780991 # Radar measured intensity in dB
Data preprocessing is done by extracting by the voxels. Voxel extraction code is available here. The file have variables which need to be set to the path of the raw data folders. The path is controlled using the below variables.
parent_dir = 'Path_to_training_or_test_data'
sub_dirs=['boxing','jack','jump','squats','walk']
extract_path = 'Path_to_put_extracted_data'
- Separate train and test files with 71.6 minutes data in train and 21.4 minutes data in test.
- Voxels of dimensions 10x32x32 (where 10 is the depth).
- Windows of 2 seconds (60 frames) having a sliding factor of 0.33 seconds (10 frames).
- We finally get 12097 samples in training and 3538 samples in testing.
- For deep learning classifiers, we use 20% of the training samples for validation.
Finally the voxels have the format: 60*10*32*32, where 60 represents the time, 10 represent the depth, and 32*32 represents the x and y dimension.
- SVM Classifier: Code
- Multi-layer Perceptron (MLP) Classifier: Code
- Bi-directional LSTM Classifier: Code
- Time-distributed CNN + Bi-directional LSTM Classifier: Code
The pretrained Bi-directional LSTM Classifier (90% accuracy) and Time-distributed CNN + Bi-directional LSTM Classifier (92% accuracy) along with the preprocessed training and test dataset are available here. The size of preprocessed dataset is around 70 GB. Classifiers code in the repo has the data loading code for the preprocessed dataset. The trained classifiers expects the test data in the same format as in the classifier training code files. The classifier can be loaded using the keras load_model function.
You can cite our paper if you have used this code in any of your projects:
@inproceedings{singh2019radhar,
title={RadHAR: Human Activity Recognition from Point Clouds Generated through a Millimeter-wave Radar},
author={Singh, Akash Deep and Sandha, Sandeep Singh and Garcia, Luis and Srivastava, Mani},
booktitle={Proceedings of the 3rd ACM Workshop on Millimeter-wave Networks and Sensing Systems},
pages={51--56},
year={2019},
organization={ACM}
}