Academic Publication (SDS)

Human activity recognition is an important functionality in any intelligent system designed to support human daily activities. While majority of human activity recognition systems use supervised learning, these systems lack the ability to detect new activities by themselves. In this paper, we report the results of our investigation of unsupervised human activity detection with features extracted from skeleton data obtained from RGBD sensor. Unlike activity recognition, activity detection does not provide the label however attempts to distinguish one activity from another. This paper demonstrates a suitable set of features to be used with K-means clustering to distinguish different activities from a pool of unlabeled observations. The results show 100% F0.5-score were achieved for six out of nine activities for one of the subjects at low frame rate, while F0.5-score of 71.9% was achieved on average for all activities by four subjects.

Human activity detection is usually achieved through supervised learning where a model is learned from given samples of each activity. With this approach, it is difficult for intelligent systems to automatically discover and learn new activities. We attempt to distinguish human activity using unsupervised learning. The approach relies on the fact that given appropriate features, the distinction between different activities can be observed from simple distance measurement. In this paper, we present our investigation to extract features from the coordinates of human joint positions based on human range of movement and the results of tests performed to check their effectiveness to distinguish sixteen (16) example activities are reported. Simple unsupervised learning, K means clustering was used to evaluate the effectiveness of the features. The results indicate that the features based on range of movement significantly improved clustering performance.

The ability to understand what humans are doing is crucial for any intelligent system to autonomously support human daily activities. Technologies to enable such ability, however, are still undeveloped due to the many challenges in human activity analysis. Among them are the difficulties in extracting human poses and motions from raw sensor data, either recorded from visual sensor or wearable sensor and the need to recognize activities not seen before using unsupervised learning. Furthermore, human activity analysis usually requires expensive sensors or sensing environment. With the availability of low-cost RGBD (RGB-depth) sensor, the new form of data can provide human posture data with high degree of confidence. In this paper, we present our approach to extract features directly from such data (joint positions) based on human range of movement and the results of tests performed to check their effectiveness to distinguish sixteen (16) example activities are reported. Simple unsupervised learning, K-means clustering was used to evaluate the effectiveness of the features. The results indicate that the features based on range of movement significantly improved clustering performance.

An approach for unsupervised human activity discovery has been proposed in this paper. The approach automatically discover unknown activities from unlabeled data and has the ability to reject random activities. This ability will enable intelligent systems to discover and learn new activities autonomously. K-means is used to cluster a pool of unlabeled activity observations into groups of different activities. The system requires no prior knowledge of how many activities to be discovered. It uses cluster validity indices to automatically estimate the required number of clusters and further evaluate cluster homogeneity to accept clusters with homogenous activity and reject clusters with random activities. Experimental results showed the potential of the approach and identified suitable validity indices to achieve unsupervised human activity discovery.

Many algorithms have been proposed to solve the clustering problem. However, most of them lack a proper strategy to maintain a good balance between exploration and exploitation to prevent premature convergence. Multi-Trial Vector-based Di erential Evolution (MTDE) is an improved di erential evolution (DE) algorithm that is done by combining three strategies and distribut- ing the population between these strategies to avoid getting stuck at a local optimum. In addition, it records inferior solutions to share information about visited regions with solutions of the next generations. In this paper, an Improved version of the Multi-Trial Vector-based Di erential Evolution (IMTDE) algorithm is proposed and adapted for clustering data. The purpose here is to enhance the balance between the exploration and exploitation mechanisms in MTDE by employing Gaussian crossover and modifying the sub-population distribution between the strategies. To evaluate the performance of the proposed clustering, 19 datasets with di erent dimensions, shapes, and sizes were employed. The obtained results of IMTDE demonstrate improve- ment in MTDE performance by an average of 12%. Our comparative study with state-of-the-art algorithms demonstrates the superiority of IMTDE in most of these datasets because of the e ective search strategies and the sharing of previous experiences in generating more promising solutions. Source code is available on Github: https://github.com/parhamhadikhani/IMTDE- Clustering.