An Efficient HOS-Based Gait Authentication of Accelerometer Data

Abstract

We propose a novel efficient and reliable gait authentication approach. It is based on the analysis of accelerometer signals using higher-order statistics. Gait patterns are obtained by transformation of acceleration data in feature space represented with higher-order cumulants. The proposed approach is able to operate on multi-channel and multi-sensor data by combining feature-level and sensor-level fusion. Evaluation of the proposed approach was performed using the largest currently available dataset OU-ISIR containing inertial data of 744 subjects. Authentication was performed by cross-comparison of gallery and probe gait patterns transformed in feature space. Additionally, the proposed approach was evaluated using dataset collected by McGill University, containing long-sequence acceleration signals of 20 subjects acquired by smartphone during casual walking. The results have shown an average EER of 6 % to 12 %, depending on the selected experimental parameters and setup. When compared to the latest state-of-the-art, evaluated performance reveal the proposed approach as one of the most efficient and reliable of the currently available accelerometerbased gait authentication approaches.

Wearable Device-Based Gait Recognition Using Angle Embedded Gait Dynamic Images and a Convolutional Neural Network

“…Similar to [27,33], we evaluate the performance with different number $\mathcal{N}$ of consecutive gait cycles as probe data, where $\mathrm{scriptN}=1,2,3,5$. The classifier was built for each subject and evaluated by 10-fold cross-validation.…”

“…The accuracy of the proposed approach is compared with the previously published approaches in Table 1 where only the results of $\mathrm{scriptN}=1,2$ are shown. To achieve the results in Table 1, [9] used a sliding window of 2.4 s and [33] choose the signal segments of length w = 2.8 s. In this context, our experimental condition was set similar to those in [9,33], because the average length of combined gait cycles when $\mathrm{scriptN}=2$ is 2.14 s.…”

“…If two contiguous gait cycles are used together as the probe data, that is $\mathrm{scriptN}=2$, which makes the probe window length similar to the length of the sliding window in [9] and [33], we can see that the proposed approach outperforms all the baseline methods in term of accuracy. The proposed approach given $\mathrm{scriptN}=2$ outperforms the Cosine Similarity method in [9] by 13.6% for the same day task, and by 5.3% for the different day task.…”

“…Note that all the feature extraction is accomplished by the CNN model directly, and the parametric model learned is compact. However, both [33] and [9] utilized nearest neighbor classifier to identify each input probe pattern. Nearest neighbor classifier is an effective non-parametric method for recognition and regression while the disadvantage lies in its low efficiency in computation and storage because most of the gallery data need to be stored to compare with the probe data.…”

“…To do this, we perform the different day testing with the same setting as illustrated above but $\mu$% of the whole dataset from data files of day 1 are selected randomly for training, similar to [33], where $\mu =10,20,\dots ,100$.…”