Window size impact in human activity recognition

Abstract

Signal segmentation is a crucial stage in the activity recognition process; however, this has been rarely and vaguely characterized so far. Windowing approaches are normally used for segmentation, but no clear consensus exists on which window size should be preferably employed. In fact, most designs normally rely on figures used in previous works, but with no strict studies that support them. Intuitively, decreasing the window size allows for a faster activity detection, as well as reduced resources and energy needs. On the contrary, large data windows are normally considered for the recognition of complex activities. In this work, we present an extensive study to fairly characterize the windowing procedure, to determine its impact within the activity recognition process and to help clarify some of the habitual assumptions made during the recognition system design. To that end, some of the most widely used activity recognition procedures are evaluated for a wide range of window sizes and activities. From the evaluation, the interval 1-2 s proves to provide the best trade-off between recognition speed and accuracy. The study, specifically intended for on-body activity recognition systems, further provides designers with a set of guidelines devised to facilitate the system definition and configuration according to the particular application requirements and target activities.

Figure 1.

Distribution of the activity recognition research studies presented in Tables 2 and 3 based on the window size.

Figure 2.

Different stages of the activity recognition chain (ARC). An example of the correlation of the windowing approach and subsequent levels of the ARC is shown. Here, different window sizes are depicted particularly. Concretely, M sensors deliver raw signals (u₁, u₂, …, u_M), which are subsequently processed (p₁, p₂, …, p_M). The signals are partitioned into data windows of size W_k (e.g., s_1Wk, s_2Wk, …, s_MWk). For each window, k, a set of features are extracted and aggregated in a single feature vector (f₁(s_1Wk), f₂(s_2Wk), …, f_M(s_MWk)) that is used as the input to a classifier. The classifier yields a class (c_Wk) that represents the identified activity.

Figure 3.

Effect of the data window size on the activity recognition system performance (F₁-score). Twelve recognition systems, respectively, corresponding to the combination of three feature sets (FS1, FS2, FS3) and four classification models (DT, NB, NCC, KNN) are evaluated.

Figure 4.

Activity-specific recognition performance for diverse window sizes and methodologies (-): (a) NCC-FS1; (b) NCC-FS2; (c) NCC-FS3; (d) NB-FS1; (e) NB-FS2; and (f) NB-FS3. The minimum window size required to achieve a specific F₁-score is depicted. No color is specified (not defined, ND) for performance values that may not be achieved for any of the window sizes and methodologies.

Figure 5.

Activity-specific recognition performance for diverse window sizes and methodologies (-): (a) DT-FS1; (b) DT-FS2; (c) DT-FS3; (d) KNN-FS1; (e) KNN-FS2; and (f) KNN-FS3. The minimum window size required to achieve a specific F₁-score is depicted. No color is specified (not defined, ND) for performance values that may not be achieved for any of the window sizes and methodologies.

Figure 6.

Minimum window size required for diverse performance thresholds. The threshold values are respectively calculated from the maximum F₁ – score that could be achieved for the recognition of each activity (represented on top). The results for two particular recognition methodologies are shown: (a) DT-FS2; and (b) KNN-FS2. Non-colored spots (not defined, ND) correspond to performance values for which no window enhancement may be obtained.