Source-linked AI summary
Deep Residual Bidir-LSTM for Human Activity Recognition Using Wearable Sensors
Yu Zhao, Rennong Yang, Guillaume Chevalier, Maoguo Gong
TL;DR
HAR involves spatial complexity and temporal divergence in behavior. The paper proposes a Deep-Res-Bidir-LSTM framework, reporting a 4.78% accuracy improvement on the public UCI data set and a 3.68% F1-score increase on the Opportunity data set compared with previous work.
Problem
HAR must address the spatial complexity and temporal divergence of human behavior while supporting its broad research value.
Method
The paper proposes a Deep-Res-Bidir-LSTM network using bidirectional LSTM connections and residual connections between stacked cells to counter gradient vanishing.
Results
4.78% higher accuracy was reported for the public UCI data set, while the Opportunity data set showed a 3.68% higher F1 score compared with previous work.
Takeaways & Limitations
The proposed Deep-Res-Bidir-LSTM framework improves reported HAR performance across the public UCI and Opportunity data sets.
Takeaways & Limitations
Future work should develop a more efficient parameter-tuning method because the grid-search range must be changed.
Abstract
from arXiv · showhide
Human activity recognition (HAR) has become a popular topic in research because of its wide application. With the development of deep learning, new ideas have appeared to address HAR problems. Here, a deep network architecture using residual bidirectional long short-term memory (LSTM) cells is proposed. The advantages of the new network include that a bidirectional connection can concatenate the positive time direction (forward state) and the negative time direction (backward state). Second, residual connections between stacked cells act as highways for gradients, which can pass underlying information directly to the upper layer, effectively avoiding the gradient vanishing problem. Generally, the proposed network shows improvements on both the temporal (using bidirectional cells) and the spatial (residual connections stacked deeply) dimensions, aiming to enhance the recognition rate. When tested with the Opportunity data set and the public domain UCI data set, the accuracy was increased by 4.78% and 3.68%, respectively, compared with previously reported results. Finally, the confusion matrix of the public domain UCI data set was analyzed.
1. Introduction
Human activity recognition addresses complex sensor-based time-series behavior, but no unified recognition method exists. The paper proposes Deep-Res-Bidir-LSTM to improve learning across temporal and spatial dimensions.
- HAR extracts motion characteristics from sensor data to recognize human activities.
- Behavioral spatial complexity and temporal divergence mean that no unified recognition method exists.
- Deep-Res-Bidir-LSTM combines bidirectional communication through time with residual connections between stacked layers.
- The network targets complex, large-scale HAR problems where sensor fusion would be required.
- Experiments compare recognition accuracy on the public domain UCI and Opportunity HAR benchmarks with other algorithms.
2. Background
The background introduces LSTM, bidirectional LSTM, and residual networks as components for modeling HAR sequences. Their combination uses temporal context and skip connections to support deeper information and gradient transmission.
- Baseline LSTM: LSTM extends recurrent neural networks with gated memory that supports longer-term time-series information.Input and forget gates control what is retained or overwritten, enabling gradients to flow through time.
- Baseline LSTM: Baseline LSTM processes sequences through recurrent hidden states and produces outputs for an upper layer.The paper defines input, output, and hidden-layer sets across time and corresponding network weight matrices.
- Bidirectional LSTM: Bidirectional LSTM combines forward and backward LSTM cells so current outputs use both previous and subsequent information.The proposed variant concatenates the two directional results and applies a fully connected layer with ReLU.
- Residual Network: Residual networks provide skip highways that transmit lower-layer information directly to upper layers.These connections are intended to preserve gradient flow while refining outputs in deeper networks.
- Residual Network: In the residual design, n = 0 corresponds to baseline deep-stacked LSTM layers without residual connections.
3. Our Model: Deep Residual Bidir-LSTM Network
The proposed Deep-Res-Bidir-LSTM combines bidirectional temporal processing with residual depth to address optimization challenges in HAR. Its pipeline preprocesses wearable-sensor time series, trains and evaluates predictions, and applies regularization and optimization techniques to control overfitting and unstable gradients.
- Data preprocessing: The HAR pipeline preprocesses wearable-sensor activity data, reshapes it into 128-time-step windows, and splits it into training and testing sets.Missing data are added, and the data are normalized to mean zero and standard deviation 0.5 before reshaping.
- Training and evaluation: Training compares predictions with real values using sigmoid cross-entropy with L2 weight decay, backpropagation, and the Adam optimizer.Accuracy and F1 score are calculated during learning and testing, while testing is added without changing learned parameters or results.
- Network architecture: Deep-Res-Bidir-LSTM combines bidirectional LSTM layers with residual connections, extending LSTM-based modeling across temporal and depth dimensions.The bidirectional structure provides forward and backward information, while residual connections are used in stacked layers.
- Limitations: The model is constrained to 128 time steps per prediction, and increasing layers or cells raises computational complexity without necessarily increasing recognition accuracy.The text states that additional depth requires regularization to avoid overfitting while improving accuracy.
- Optimization rationale: Residual connections act as gradient highways that help counter gradient vanishing and make otherwise difficult-to-optimize lower layers more accessible during backpropagation.Batch normalization is placed above each highway layer in the described architecture.
- Network architecture: The described network contains two residual hidden layers, each with two bidirectional layers, corresponding to eight LSTM cells in total.The architecture uses ReLU as its activity function.
4. Experiments
Experiments evaluated Deep-Res-Bidir-LSTM on the public domain UCI and Opportunity datasets using dataset-specific preprocessing, splits, and metrics. The model outperformed comparison methods, with strong UCI classification results and improved Opportunity gesture recognition.
- Experimental setup: Experiments used the public domain UCI and Opportunity datasets, comparing Deep-Res-Bidir-LSTM against other recognition methods.The UCI experiments used 30 volunteers performing six activities, while Opportunity experiments used wearable, object, and ambient sensor data.
- UCI dataset: The UCI data were collected at 50 Hz, windowed into 2.56-second segments, and divided into volunteer-based training and test sets.Each window contained 128 time steps, and 70% of volunteers were used for training while 30% were used for testing.
- Results: 93.54% accuracy was achieved on the UCI dataset, with residual connections and bidirectional cells identified as contributing factors.The model’s accuracy and F1 columns were reported as nearly identical, and its results outperformed the comparison algorithms.
- Results: 3.68% better gesture recognition was reported for the Opportunity dataset, where Deep-Res-Bidir-LSTM achieved the best F1 score at approximately 0.9.The model converged faster than baseline LSTM and Bidir-LSTM, while its convergence was close to Res-LSTM.
- Confusion-matrix analysis: 93.57% integral accuracy was obtained on UCI test data, with strongest recognition for LAYING_DOWN and confusion between static and dynamic activity classes.Standing and sitting were sometimes confused, as were WALKING, WALKING_UPSTAIRS, and WALKING_DOWNSTAIRS.
5. Conclusions
The Deep-Res-Bidir-LSTM framework improves HAR learning and information transmission through residual connections and bidirectional cells. Experiments report higher benchmark performance, while conclusions identify hyperparameter tuning and generalization as important boundaries for future work.
- Conclusions: The proposed Deep-Res-Bidir-LSTM framework enhances learning ability and supports information transmission through residual and bidirectional connections.Residual connections operate on the depth dimension, while bidirectional cells operate on the temporal dimension.
- Results: 4.78% accuracy improvement was reported for the public domain UCI data set compared with previous work.
- Results: 3.68% F1 score improvement was reported for the Opportunity data set compared with previous work.
- Hyperparameter findings: Window size was a key parameter: too-small windows disrupted information continuity, whereas too-large windows caused classification errors.A window size from 500 ms to 5000 ms was described as usually appropriate.
- Hyperparameter findings: The suitable number of cells differed by data set: 28 cells were sufficient for UCI, while 128 were better for Opportunity because of its greater complexity.
- Future work: Future work should develop adaptive parameter search and architecture evolution, including automatically reshaping, adding, or removing layers.The paper also proposes exploring mixed 1D time-based convolutions within LSTM cells and applying the network to other fields.