PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Recurrent Neural Network for Mobile Human Activity Recognition with High Throughput

Masaya Inoue, Sozo Inoue, Takeshi Nishida

arXiv:1611.03607v1cs.CVcs.NE

TL;DR

Mobile human activity recognition needs accurate, fast processing from sensor data. The paper develops and tunes a deep recurrent neural network using raw accelerometer data, achieving higher recognition rates than traditional methods and shorter recognition time. The authors also evaluate tuned parameters on an additional dataset and discuss training-time and sequential-transition constraints.

  • Problem

    Mobile activity recognition must support applicable care-related domains while avoiding the feature extraction and processing costs associated with traditional or CNN-based approaches.

  • Method

    The paper applies a deep recurrent neural network to raw accelerometer data and investigates architectures and parameter combinations using activity-recognition datasets.

  • Results

    95.42% and 83.43% recognition rates were achieved on single-activity and multiple-sequential test data, versus 71.65% and 54.97% for traditional methods; recognition took 1.347 [ms] versus 11.031 [ms].

  • Takeaways & Limitations

    The tuned DRNN provided higher recognition rates and shorter recognition time than traditional methods, and achieved 95.03% recognition on an additional HAR dataset.

  • Takeaways & Limitations

    Training averaged 116.39 [s] per epoch, although the authors suggest precomputed training and mini-batch online learning could support high-throughput training.

Abstract

from arXiv · show

In this paper, we propose a method of human activity recognition with high throughput from raw accelerometer data applying a deep recurrent neural network (DRNN), and investigate various architectures and its combination to find the best parameter values. The "high throughput" refers to short time at a time of recognition. We investigated various parameters and architectures of the DRNN by using the training dataset of 432 trials with 6 activity classes from 7 people. The maximum recognition rate was 95.42% and 83.43% against the test data of 108 segmented trials each of which has single activity class and 18 multiple sequential trials, respectively. Here, the maximum recognition rates by traditional methods were 71.65% and 54.97% for each. In addition, the efficiency of the found parameters was evaluated by using additional dataset. Further, as for throughput of the recognition per unit time, the constructed DRNN was requiring only 1.347 [ms], while the best traditional method required 11.031 [ms] which includes 11.027 [ms] for feature calculation. These advantages are caused by the compact and small architecture of the constructed real time oriented DRNN.

1. INTRODUCTION

Human activity recognition supports care-related applications, while mobile sensors have enabled active research in low-cost sensing. This paper applies a recurrent neural network to raw accelerometer data and evaluates accuracy and recognition speed against traditional methods.

  • Human activity recognition applies to health care, preventive medicine, and elderly care, motivating mobile sensing research.
  • Deep learning can directly process original data, potentially avoiding feature-vector calculation during training and recognition.
  • The proposed method recognizes raw accelerometer data with a recurrent neural network while investigating architectures and parameter combinations.
  • 95.42% recognition rate was achieved on 108 single-activity test trials, compared with 71.65% for the traditional method.The model was trained on 432 segmented trials covering 6 activity classes from 7 people.
  • 83.43% recognition rate was achieved on 18 multiple sequential trials, compared with 54.97% for the traditional method.

1. In order to construction of a fast response classifier oriented

The study develops a real-time-oriented recurrent neural network and explores parameters affecting accuracy, then evaluates recognition throughput against traditional methods.

  • The study adopts a recurrent neural network architecture for real-time execution and evaluates its advantages against traditional methods.
  • Various parameters were explored to investigate factors affecting recurrent-neural-network accuracy using two datasets.

2. BACKGROUND AND RELATED WORK

Traditional activity recognition commonly extracts feature vectors for machine-learning classifiers, whereas deep CNN and RNN methods can use raw sequential signals. The paper focuses on a deep RNN intended to combine raw-data processing, accuracy, and high throughput for mobile recognition.

  • Traditional methods commonly classify feature vectors extracted from signal windows using decision trees, k-nearest neighborhood, naive Bayes, support vector machines, or random forests.
  • Sequential-data recognition has also used hidden Markov models and conditional random fields, while real-time studies often reduced feature-calculation costs.
  • Deep CNN methods can improve accuracy over traditional methods but require fixed-length time-window segmentation and generally have many inner-layer connections unsuitable for mobile real-time execution.
  • RNNs are suitable for time-series data, and deep learning treats an RNN as having multiple temporal-direction layers.
  • Combining CNN and RNN can increase recognition accuracy, but CNN adoption also increases computational cost and requires a time window.
  • The proposed deep RNN trains and recognizes raw acceleration data without prior feature extraction, aiming for high-precision recognition with high throughput.

3. RECURRENT NEURAL NETWORK

The paper introduces the basic processing methods used for recurrent-neural-network training and recognition.

  • The section explains the basic processing methods for executing recurrent-neural-network training and recognition.

3.1 Deep recurrent neural model

The paper defines a deep recurrent neural network with recurrent connections within each internal layer and direct raw-data-to-output processing. Layer outputs are transformed by activation functions to produce the final network output.

  • The DRNN is an L-layer Elman-type network with fully connected recurrent connections within each hierarchy over time.Each layer receives input from the preceding layer and recurrent input from the previous time step.
  • The first-layer input is the raw vector x_k, while the final layer produces output y_k from v_k.The output layer has no recurrent connection, so its output depends on the current final-layer input.
  • Each layer computes unit outputs by applying an activation function to weighted inputs and recurrent values.The paper lists sigmoid, logistic sigmoid, and ReLU as frequently used activation functions.
  • Biases are incorporated by introducing a 0-th weight and a constant 0-th unit.This allows biases to be represented collectively in the layer computation.
  • Figure 1 schematically represents the DRNN architecture.

3.2 Learning method

The learning method uses softmax classification with cross-entropy error, optimized through gradient-based updates. Back propagation through time trains recurrent connections, while truncated BPTT limits computational cost.

  • Softmax outputs represent class-membership probabilities, and the input is assigned to the class with the largest probability.
  • The cross-entropy function compares supervised class vectors with predicted probabilities and is minimized during learning.The target vector uses 1 for the correct class and 0 otherwise.
  • Gradient descent updates the parameter vector by repeatedly moving it in the negative gradient direction using a learning rate.The learning rate controls the update magnitude, and Adam is used when adjustment is needed.
  • Mini-batch stochastic gradient descent computes gradients from the average error over a small batch of samples.The paper states that this method converges quickly and has high local-solution-avoidance performance.
  • BPTT propagates deltas backward through both layers and time, while truncated BPTT restricts the time span to a fixed constant for practical computation.

3.3 Long short-term memory (LSTM)

LSTM augments recurrent networks with memory-cell and gating mechanisms for controlling stored information. Its structure addresses vanishing gradients, input/output weight conflicts, and abrupt changes in time-series patterns.

  • LSTM uses a memory cell to control whether information is written, read, or deleted.
  • The Constant Error Carousel preserves the internal state across time and addresses the vanishing gradient problem during recurrent back-propagation.
  • Input and output gates regulate how much information passes and whether memorized state is read or retained.The logistic function produces values from 0 to 1, providing the gate-control range.
  • The forget gate determines whether the memorized state is forgotten, supporting learning when the time-series pattern changes suddenly.
  • LSTM increases the number of weights by adding gate-specific transmission and bias weights.
  • Gradient clipping prevents the gradient’s L2 norm from exceeding a threshold, addressing exploding gradients.

3.4 Avoiding overfitting

The paper discusses overfitting as a risk when network freedom exceeds the available training data. It considers regularization and dropout as ways to reduce this risk while retaining model expressiveness.

  • Overfitting occurs when a model captures peculiarities of training data and fails to fit new test data.
  • Overfitting is likely when network degrees of freedom are too high for the training data.
  • Regularization constrains learning parameters to mitigate overfitting when additional training data are difficult to obtain.The paper also identifies gradient clipping as a regularization technique.
  • Dropout disables internal and output units at a fixed rate during training, while all units are used during recognition.
  • The paper motivates examining layer count, unit count, truncated time, and dropout rate because RNN performance depends on parameter choices.

4. ACTIVITY RECOGNITION USING RNN

The study evaluates a DRNN for human activity recognition using HASC data, comparing accuracy, parameter effects, and recognition throughput with traditional methods.

  • Dataset and evaluation: The evaluation uses HASC segmented and sequence data, with 432 training trials and 108 segmented test trials.Segmented trials contain single activities, whereas sequence trials contain multiple consecutive activities.
  • Dataset and evaluation: Accuracy is measured as the proportion of correctly recognized samples, separately for training, segmented test, and sequence data.Training, test, and sequence accuracy use different recognition datasets after training on the training data.
  • DRNN architecture and training: The DRNN maps three-axis acceleration at each time to a three-dimensional input and six activity classes to a six-dimensional output using LSTM internal units.The output uses softmax activation and cross-entropy loss, with truncated BPTT and mini-batch stochastic gradient descent.
  • DRNN architecture and training: Training shuffles trials into mini-batches, samples a starting time, applies truncated BPTT across time ranges, and updates weights through back-propagation.The final time range is bounded by K′ = 1200.
  • Comparative methods: Decision tree, SVM, and random forest baselines use feature vectors computed from five-second windows shifted every 2.5 seconds.The baselines use selected features after reducing 27 computed variables to 13 through stepwise feature selection.
  • Throughput evaluation: Throughput is computed per sample, while baseline timing separately includes feature calculation and activity recognition.The DRNN is evaluated on a CPU, whereas GPU parallel processing reduces training time.

5. RESULTS

The DRNN achieves strong accuracy across segmented and sequential activity data, with performance depending on several architectural and training parameters and with substantially faster overall recognition than feature-based baselines.

  • Best model: 95.42% was the best test recognition rate, while sequence-data recognition reached 83.43%.The best model was selected by its sequence-data recognition result.
  • Comparison with existing methods: 35.18%, 27.76%, and 22.35% were the DRNN’s test-data improvements over decision tree, SVM, and random forest, respectively.For sequence data, the corresponding improvements were 28.03%, 26.04%, and 26.74%.
  • Parameter variation: Three internal layers produced the highest recognition rate across training, test, and sequence data.Sequence recognition was about 8.2% higher than with the worst four-layer model.
  • Parameter variation: 60 internal-layer units gave the highest test and sequence recognition rates among models using 20, 40, 60, and 80 units.Sequence recognition was about 3.7% higher than with the lowest 20-unit model.
  • Parameter variation: T = 70 was among the relatively strong truncated-time settings, whereas T = 10 was worst, producing a sequence-data difference of about 10.7%.The reported relatively good settings were T = 30 and T = 70.
  • Additional dataset: 95.03% recognition was obtained on the HAR dataset at the 45th epoch using the parameters found on HASC.Evaluation used the first three-fourths of samples for training and the last one-fourth for testing.

6. DISCUSSION

The best DRNN achieved strong recognition accuracy and fast inference, while performance depended on sequence complexity and architecture choices. The study also identifies training-time and future-method limitations.

  • 6.1 Activity recognition with the best model: 95.42% recognition rate was obtained for segmented test data, while sequence-data recognition reached 83.43% at maximum.The authors prioritize test and sequence-data accuracy over training accuracy as the basic evaluation criterion.
  • 6.1 Activity recognition with the best model: Similar activities were erroneously recognized around time 6000–7000, and recognition was delayed near time 2300 in sequence data.The delay was attributed to unlearned transition signals that could not be classified into an activity.
  • 6.2 Parameter examination: Three internal layers produced 116.39 [s] per epoch and 1.347 [ms] recognition throughput, whereas four layers increased these costs to 144.83 [s] and 1.720 [ms].Training accuracy increased through three internal layers but decreased in the four-layer model, while computation time and memory usage increased with depth.
  • 6.2.2 Truncated time: Recognition performance decreased at truncated time T = 10, while relatively high recognition rates occurred at T = 30 and T = 70.The authors relate truncated-time selection to the variation period of the handled signals.
  • 6.2 Parameter examination: Dropout tended to improve recognition rates for test and sequence data, with approximately 50% considered appropriate for the RNN.Gradient-clipping parameters produced no significant performance difference in this experiment, although suppression becomes more relevant with longer truncation and deeper networks.
  • 6.3 Throughput and training time: 1.347 [ms] recognition throughput was 8.19 times faster than the existing method and was considered sufficient for 100 [Hz] real-time processing.The compact DRNN had 74,166 internal variables and architectural elements, less than 10% of the conventional CNN+LSTM model.
  • 6.3 Throughput and training time: 116.39 [s] average training time per epoch remained large and non-negligible, although precomputed high-speed training may preserve high-throughput recognition.The authors also identify verification of newer RNN techniques and approaches for sequence-data recognition as future work.

7. CONCLUSION

The paper constructs a compact DRNN for mobile-device activity recognition from raw acceleration time series. It reports high recognition rates and substantially faster throughput than traditional methods, while identifying further methods for future study.

  • 7. CONCLUSION: 95.42% and 83.43% maximum recognition rates were achieved on segmented and multiple-sequential test datasets, respectively, versus 71.65% and 54.97% for traditional methods.The tuned parameters were also evaluated using an additional sequential dataset.
  • 7. CONCLUSION: 1.347 [ms] recognition throughput was achieved by the DRNN, compared with 11.031 [ms] for the traditional method, including 11.027 [ms] for feature extraction.The paper attributes these advantages to the DRNN’s compact architecture.
Loading 1611.03607v1…