Statistical Noise Reduction for Robust Human Activity Recognition Song-Mi Lee, Heeryon Cho, and Sang Min Yoon HCI Lab., College of Computer Science, Kookmin University, Seoul, South Korea
[email protected],
[email protected]
SUMMARY
① Human Activity Sensing x y z
Noise and variability in accelerometer data collected using smart devices obscure accurate human activity recognition
a x2 y2 z 2
We apply a statistical noise reduction using total variation minimization to accelerometer data Experimental results using Random Forest classifier prove that our noise removal approach is effective
④ Learning & Classification
③ Noise Reduction
Run Walk
Combining our approach with 1D-CNN, however, was counterproductive
OUR APPROACH
② Feature Vector Transformation
Still
Acceleration values at 𝑥, 𝑦, 𝑧 axes are transformed into a vector magnitude by calculating the Euclidean norm at time 𝑡: 𝑢 𝑡 =
𝑥(𝑡)2 + 𝑦(𝑡)2 + 𝑧(𝑡)2
Suppose that the observed data 𝑢(𝑡) is a discrete sampling from the accelerometer data, and the signal of time length 𝑁 is contaminated by noise Magnitude data can be written as 𝑢 𝑡𝑛 = 𝑣 𝑡𝑛 + 𝜖𝑛 , 𝑛 = 1, … , 𝑁, where 𝑣(𝑡𝑛 ) is the value of an underlying function 𝑣, and 𝜖𝑛 indicates the additive noise at time 𝑡 From a given noisy input, i.e., discrete signal 𝑢(𝑡𝑛 ), defined on 𝑇 = {𝑡1 , 𝑡2 , … , 𝑡𝑁 }, the Total Variation (TV) minimization technique generates an estimated signal 𝑣 which retains the movement of the data and removes the noisy input signal 𝑢 by solving the following minimization problem: Fig. Vector magnitude (top) and noise-reduced (bottom, λ=6) signals
EVALUATION ON COLLECTED DATASET Evaluated 3-class activity recognition (run, walk, still)
Compared random forest & 1d-CNN classifiers’ accuracy on raw vs. noiseremoved (λ=6) data Random forest classifier built on TV minimization-applied data performed the best Feature learning approaches such as CNN performed worse when TVminimization was applied to data
𝑁
(𝑤 𝑡𝑛 − 𝑢 𝑡𝑛 )2 +𝜆
𝑣 = argmin 𝑤:𝑇→ℝ
𝑁−1
𝑛=1
𝑤 𝑡𝑛+1 − 𝑤(𝑡𝑛 ) 𝑛=1
ON OPEN DATASET Evaluated 7-class activity recognition (static, moving slowly,
walking, running, driving a car, riding a bus, riding a train)
Compared random forest classifier performance using raw vs. noise-removed (λ=1) data 30 trials of 10-fold CV, mean accuracy was measured Raw data vs. noise-removed data= 85.4% vs. 85.7% Wilcoxon signed-rank test showed that accuracy difference was significant (𝑝 < 0.008)
This work was supported by the Institute for Information and Communications Technology Promotion (IITP) grants (No. 2014-0-00501 & No. 2017-0-00205) and by the National Research Foundation of Korea (NRF) funded by the Ministry of Education (NRF-2016R1D1A1B04932889, NRF-2015R1A5A7037615) and the Ministry of Science and ICT of Korea (NRF-2017R1A2B4011015).
