EMG-based Gesture Recognition using Capsule Network

Data

The NinaPro dataset will be downloaded automatically by running the main.py script if it is not found.

Requirements

1. python 3.7.4
2. numpy 1.18.1
3. scikit-image 0.16.2
4. torch 1.4.0
5. tqdm 4.44.1
6. matplotlib 3.1.3

TODO

• Build dataset downloader and loader
• Build feature extractors
• Build general classification engine
• Build general training/evaluation engine
• Build conventional classifiers (Support Vector Machine, Random Forest)
• Build Neural Network based classifiers (Fully Connected Network, Convolutional Neural Network)
• Build Capsule Network based classifier
• Create logging and plotting engines
• Train and Evaluate

Train

Multiple data types, feature types and models could be chosen. For instance, the following command line runs a training of a capsule net using the base dataset and the root mean square features.

python train.py --model capsnet --features rms --data intrasubjects --epochs 50 --chkpt_period 1 --valid_period 1 --batch_size 5 --lr 0.001 --verbose 1

Run python train.py -h to see the possible values of the arguments.

Results

Testing on Intrasubject Fixed Window Dataset

Features	rf	SVM	FCN	CNN	EMGCaps
rms	66.56%	52.06%	67.67%	55.54%	76.04%
hist	56.34%	38.58%	49.28%	53.88%	65.67%
multirms	65.29%	50.93%	57.26%	61.47%	77.85%
pmrms	66.44%	63.15%	63.29%	61.72%	76.41%
kmrms	54.4%	25.57%	38.73%	39.65%	67.47%
fourier	62.32%	48.34%	57.14%	56.77%	69.18%

Testing on Intrasubject Variable Window Dataset

Features	rf	SVM	FCN	CNN	EMGCaps
rms	79.53%	53.27%	74.67%	66.26%	90.56%
multirms	50.93%	66.82%	48.69%	81.21%	93.27%
pmrms	81.96%	82.42%	48.69%	81.21%	92.52%
kmrms	64.67%	23.92%	15.88%	39.65%	85.04%

Testing on Intersubject Fixed Window Dataset

Features	rf	SVM	FCN	CNN	EMGCaps
rms	15%	14.8%	17.42%	16.2%	15.56%
hist	14.82%	14.95%	14.67%	14.88%	15.21%
multirms	15.97%	15.03%	15.51%	15.49%	17.57%
pmrms	9.8%	13.4%	11.93%	10.97%	11.78%
kmrms	11.85%	11.07%	14.09%	13.4%	15.00%
fourier	16.02%	14.95%	16.98%	17.42%	20.54%

Testing on Intersubject Variable Window Dataset

Features	rf	SVM	FCN	CNN	EMGCaps
rms	15.2%	15.67%	19.27%	19.59%	20.84%
multirms	22.88%	20.53%	27.27%	27.11%	28.99%
pmrms	18.18%	23.19%	20.68%	20.06%	23.51%
kmrms	14.89%	10.81%	12.06%	18.02%	19.74%