Enhanced bag-of-words representation for human activity recognition using mobile sensor data

Human activity recognition based on sensors (e.g., accelerometer and gyroscope) embedded in smartphones is of great significance for many applications under uncontrolled environments. Although significant progress has been noticed in this field, one of the challenges limiting its real-life applications lies in robust feature extraction for efficient activity recognition on smartphones. This study addresses this challenge by proposing an improved bag-of-words representation for activity signal characterization. Specifically, raw activity signals are processed by discrete wavelet transformation to extract local features, which will be clustered using K-means to form a bag-of-words dictionary. The vocabularies in the dictionary are regarded as bin centers for histogram feature construction. For each local feature of an activity signal, its distance from all the bin centers will be measured. To capture higher-order information for feature representation, the frequency for the bin centers corresponding to the minimum n distances will be updated. Moreover, the frequency is increased by a trigonometry constraint cosine value of the corresponding distances to account for activity signals' structural information. The proposed feature representation has been verified with three well-established classifiers, namely SVM, ANN, and KNN on the UCI-HAR dataset. The consistently good performance validates the effectiveness and robustness of the proposed feature representation. Compared with the state-of-the-art, the experimental results also demonstrate the advantage of the proposed method in terms of accuracy and computational cost.