Multiwindow Fusion for Wearable Activity Recognition
Download as PPTX, PDF0 likes299 views
The document presents a novel multiwindow fusion technique for wearable activity recognition that utilizes multiple window sizes to enhance recognition accuracy. It employs a weighted decision fusion mechanism to capitalize on the strengths of various recognition systems and demonstrates significant performance improvements over traditional single-window approaches. The experimental setup involves diverse sensor data and multiple machine learning techniques, validating the proposed model's capability to effectively improve recognition outcomes.
Multiwindow Fusion for Wearable Activity Recognition
- 1. Multiwindow Fusion for Wearable Activity Recognition Oresti Baños, Choong Seon Hong, Sungyoung Lee Department of Computer Engineering Kyung Hee University, Korea {oresti,cshong,sylee}@khu.ac.kr Juan Manuel Gálvez, Miguel Damas, Alberto Guillén, Luis Javier Herrera, Héctor Pomares, Ignacio Rojas, Claudia Villalonga Department of Computer Architecture and Computer Technology University of Granada {jonas,mdamas,aguillen,jherrera,hector,irojas,cvillalonga}@ugr.es The recognition of human activity has been extensively investigated in the last decades. Typically, wearable sensors are used to register body motion signals that are analyzed by following a set of signal processing and machine learning steps to recognize the activity performed by the user. One of the most important steps refers to the signal segmentation, which is mainly performed through windowing approaches. In fact, it has been proved that the choice of window size directly conditions the performance of the recognition system. Thus, instead of limiting to a specific window configuration, this work proposes the use of multiple recognition systems operating on multiple window sizes. The suggested model employs a weighted decision fusion mechanism to fairly leverage the potential yielded by each recognition system based on the target activity set. This novel technique is benchmarked on a well-known activity recognition dataset. The obtained results show a significant improvement in terms of performance with respect to common systems operating on a single window size. Keywords: Activity recognition, Segmentation, Windowing, Wearable sensors, Ensemble methods, Data fusion Abstract • Impact of the segmentation phase on the accuracy of the recognition models through existing relation among activity categories and involved body parts with the window size utilized • Fusion mechanisms to benefit from the utilization of several window sizes instead of restricting to a single one • Innovative multiwindow fusion technique: Weight and combine the decisions provided by multiple activity recognizers configured to operate on different window sizes of the same input data To the use of a decision fusion mechanism • Use of various levels of segmentation: two-step fusion process 1. The decisions provided by each individual activity recognizer are locally weighted and aggregated to yield a sole recognized activity per chain 2. The activities identified for each chain are combined to eventually deliver a unique recognized activity • General proposed model with N classes or activities (n = 1,…,N), set of raw (u) or preprocessed (p) sensor data that are segmented with Q different window sizes (W1,…, WQ-1, WQ) divisors of WQ (that represents the system recognition period). Q independent recognition chains and data window of size WQ (i.e. sWQ) are created, split into WQ/Wk segments of size Wk. Each segment si Wk is transformed into features f(si Wk) which are input to each respective classifier, yielding a recognized activity or class ci Wk 1. First Fusion Level: Weighted & Averaged of each individual classifier decisions (all segments and all classes) 2. Predicted class for each classifier after the First Fusion Level 3. Second Fusion Level: Weighted & Averaged of decisions obtained for each respective window size 4. Eventual recognized class Multiwindow Fusion Foundations STRUCTURAL SCHEMA (Example for Q = 5 different window sizes) EXPERIMENTAL SETUP • Dataset with motion data, namely, acceleration, rate of turn and magnetic field orientation • 17 volunteers / 33 fitness activities / 9 inertial & attached sensors • 5 window sizes considered (W1 = 0.25 / W2 = 0.75 / W3 = 1.5 / W4 = 3s / W5 = 6), all in seconds • Mean and standard deviation for the feature extraction stage • 4 machine learning techniques for the classification stage: C4.5 decision trees (DT), k-nearest neighbors (KNN), naive Bayes (NB) and nearest centroid classifier (NCC) EVALUATION & CONCLUSIONS • Improve the recognition capabilities across the use of multiple window sizes • Different number of windows required for improving the performance • Higher computation complexity with multiple windows considered • Reduce the recognition time Applying the Multiwindow Fusion This work was partially supported by the Industrial Core Technology Development Program, funded by the Korean Ministry of Trade, Industry and Energy, under grant number #10049079. This work was also funded by the Junta de Andalucia Project P12-TIC-2082. Acknowledgements