ONE-AGAINST-ALL BINARIZATION CLASSIFICATION STRATEGY TO RECOGNIZE INTERCLASS SIMILARITIES ACTIVITIES FROM SEVERAL SENSOR POSITIONS

Abstract

Prior knowledge in pervasive computing recently has garnered a great deal of attention due to the high demand in most applications in order to fulfil the human needs. Human Activity Recognition (HAR) has considered every bit unitary of the applications that are widely explored to provide the valuable information to the human. Small in size within the various smartphones, accelerometer sensor has utilized to undergo the HAR research. Current HAR is not only covered the simple daily activities but also, broadly covered the complex activities. Nevertheless, the existence of high interclass similarities activities tends to increase the level of incorrectly classified instances. Hence, this study demonstrates the binarization classification strategy to tackle the abovementioned issue for the activities with a high degree of similarities. Acceleration signal in the time domain is transformed into frequency terms for separating the signals between gravitational and body acceleration. Two different groups of features; statistical, and frequency analysis are extracted in order to increase the diversity in differentiating between stationary and locomotion activities. The problem complexity is simplified using the binarization strategy before the extracted subset is evaluated. One-Against-All (OAA) classification strategy is introduced to tackle the challenge in improving the accuracy for very similar activity. The proposed work significantly resulted with high accuracy performance, particularly in differentiating between the various high interclass similarities activities using two physical activity datasets; WISDM and PSRG.

Keywords

• Accelerometer
• Binarization
• HAR
• Interclass similarities
• Oneagainst-all
• WISDM