Source-linked AI summary
A Lightweight Deep Learning Model for Human Activity Recognition on Edge Devices
Preeti Agarwal, Mansaf Alam
TL;DR
The paper addresses the difficulty of deploying computationally demanding deep-learning HAR models on resource-constrained edge devices. It proposes a lightweight RNN-LSTM model using fixed accelerometer windows and evaluates it on six activities. The model achieves 95.78% accuracy and outperforms many previously developed machine-learning and deep-learning models, while remaining scoped to a single tri-axial accelerometer and static windowing.
Problem
Deep-learning HAR models can require substantial computation, making deployment on resource-constrained edge devices difficult despite the benefits of edge-based activity recognition.
Method
The paper develops a shallow RNN-LSTM model with two hidden layers and 30 neurons that processes fixed-size accelerometer windows on edge devices.
Results
The model achieves 95.78% accuracy, 95.81% precision, 95.78% recall, and 95.73% F1-score, and outperforms many existing machine-learning and deep-learning models in accuracy.
Takeaways & Limitations
The results support using the lightweight model for six-activity HAR on resource-constrained edge hardware such as Raspberry Pi3.
Takeaways & Limitations
The system uses static sliding windows and a single tri-axial accelerometer, with extensions to dynamic windowing and multisensor data left for future work.
Abstract
from arXiv · showhide
Human Activity Recognition (HAR) using wearable and mobile sensors has gained momentum in last few years, in various fields, such as, healthcare, surveillance, education, entertainment. Nowadays, Edge Computing has emerged to reduce communication latency and network traffic.Edge devices are resource constrained devices and cannot support high computation. In literature, various models have been developed for HAR. In recent years, deep learning algorithms have shown high performance in HAR, but these algorithms require lot of computation making them inefficient to be deployed on edge devices. This paper, proposes a Lightweight Deep Learning Model for HAR requiring less computational power, making it suitable to be deployed on edge devices. The performance of proposed model is tested on the participants six daily activities data. Results show that the proposed model outperforms many of the existing machine learning and deep learning techniques.
1. Introduction
HAR uses wearable and smartphone sensor data to identify daily activities, while edge processing can reduce the latency, cost, and network traffic associated with cloud transmission. The paper proposes and evaluates a lightweight deep-learning model for this setting.
- Continuous transmission of accelerometer signals to the cloud can increase network traffic and latency, whereas edge-based identification can reduce communication latency, cost, network traffic, and response time.
- The paper proposes a lightweight deep-learning HAR model designed for deployment on edge devices.
- The model combines a shallow recurrent neural network with long short-term memory for HAR.
- The study trains and tests the model on six activities using a resource-constrained Raspberry Pi3 and the WISDM dataset from 29 participants.
- Performance is evaluated with accuracy, precision, recall, F-measure, and confusion matrices, then compared with previously developed models.
2. Background and Related Work
HAR pipelines typically collect, preprocess, extract features, classify activities, and evaluate predictions. Deep learning can automate feature extraction and model temporal information, but computational demands motivate a lightweight RNN-LSTM approach for edge devices.
- A basic HAR system proceeds through data collection, preprocessing, feature extraction, activity classification, and evaluation.
- Preprocessing cleans sensor signals and segments them into windows using approaches such as sliding-window, event-based, or energy-based segmentation.
- Figure 1 summarizes the basic sequence of stages in a HAR system.
- Deep-learning methods are used in HAR because they can automatically extract features, but many require high computation power that complicates edge deployment.
- RNN-LSTM models address temporal correlations in sensor sequences that traditional classifiers miss and convolutional kernels restrict.
3. Description of Proposed Lightweight Model
The proposed model combines recurrent processing with LSTM memory cells in a shallow architecture intended for resource-constrained edge devices. Fixed accelerometer windows pass through two lightweight hidden layers, whose outputs are combined into one predicted activity.
- The model combines RNN and LSTM processing in a shallow structure with two hidden layers and 30 neurons, supporting deployment on constrained edge devices.
- RNNs capture temporal information from sequential sensor data through recurrent hidden states and outputs.
- LSTM cells replace traditional recurrent nodes to model long activity windows while addressing exploding and vanishing gradients through gated memory.
- 3.3. Proposed Lightweight RNN-LSTM Model: The system partitions accelerometer readings into fixed-size windows and feeds each sequence into the lightweight RNN-LSTM model.
- 3.3. Proposed Lightweight RNN-LSTM Model: Outputs from different states are combined using a sum rule and softmax classifier to produce one activity output for each window.
- 3.3. Proposed Lightweight RNN-LSTM Model: The architecture uses discrete-time inputs across two hidden layers, with the final output representing the predicted activity class for the window.
4. Experiment
The experiment evaluates the lightweight model on tri-axial smartphone accelerometer data from six activities and trains it on a Raspberry Pi3. It uses a 70:30 train-test split and standard classification metrics.
- The experiment runs on a Raspberry Pi3 with a 1.2 GHz quad-core ARM processor and 1 GB RAM.
- The WISDM data contain six activities performed by 29 subjects using a smartphone accelerometer sampled at 20 Hz.
- The dataset contains 1,098,207 total samples.
- The model is trained and tested using a 70:30 dataset split, mean cross entropy, and the Adam optimizer.
- Performance is measured with accuracy, precision, recall, and F1-score, using true- and false-positive and negative predictions.
5. Result
The Lightweight RNN-LSTM model achieved strong overall evaluation metrics and outperformed many existing machine learning and deep learning methods. Activity-level accuracy ranged from 81% for upstairs to 99% for jogging and walking.
- 99% accuracy was achieved for jogging and walking, while upstairs had the minimum accuracy of 81%.
- 95.78% accuracy, 95.81% precision, 95.78% recall, and 95.73% F1-score were achieved overall.
- The model outperformed the compared CNN, RNN, DBN, SVM, PCA, hybrid, fusion, Hidden Markov, ensemble, random forest, and multilayer perceptron methods in accuracy.
6. Conclusions and Future Directions
The paper concludes that its Lightweight HAR model can be deployed on a Raspberry Pi3 and achieves better results than many existing models. Future work targets broader activities, additional devices, dynamic windowing, and multisensor inputs.
- The Lightweight HAR model was deployed on a Raspberry Pi3 and produced better results than many existing machine learning and deep learning models.
- Future work includes recognizing more complex activities and deploying the model on Android and iOS devices.
- The current system uses static sliding windows and a single tri-axial accelerometer, motivating future dynamic-window and multisensor extensions.