**1. Introduction**

Knowledge of context, with respect to the activity performed by a user, promotes the efficiency of human-centric technologies. Especially in an indoor setting, human activity recognition is beneficial for applications such as personalised health monitoring, building energy management, security and safety. Most activity recognition approaches use custom devices in order to gather data related to the activity performed. As we discuss in Section 2, these specialised devices are either worn on multiple body parts or are installed in various locations inside the building, forming a wireless sensor network. This can include a network of pressure, temperature, humidity and acoustic sensors installed in the area [1], proprietary sensors attached to objects within specific areas [2,3], optical monition capturing systems [4] and RFID tags [5].

These approaches, however, are obtrusive and suffer in terms of practicality, as multiple specialised devices have to be installed on various objects and in different locations inside the area. This also affects integration and user acceptance, as most of the times, these devices use communication protocols (e.g., ZigBee) that are not compatible with devices such as a mobile phone carried by a typical user. The goal of this research is to accurately recognise activities related to specific areas in an indoor space by only using commercial off-the-shelf devices and investigate the effect that information related to the user's location has on the system's performance.

To achieve this, we have designed and developed a system that is composed of a smart watch, BLE beacons, a mobile phone and a server. The system collects and processes data coming from these devices, without relying on specific or customised implementations, which results in enhanced flexibility. The popularity of wearable devices has significantly increased in recent years [6], while BLE beacons have become extremely popular, and there is a wide range of commercial offerings available from multiple manufacturers [7]. In previous work [8], we investigated the feasibility of activity recognition using commercial smart watches. Here, BLE beacons are used to enhance our activity recognition system in an unobtrusive way with information regarding the location of the occupants. In particular, in this work, we investigate our system's performance when we fuse the inertial data coming from a commercial smart watch with data coming from BLE beacons. We have also evaluated different classification algorithms, feature types and segmentation window sizes. Our evaluation is based on real-world experiments, using our proposed system, that took place in an indoor laboratory environment.

In particular, the first contribution of this work is the development of an activity recognition system that incorporates commercial off-the-shelf BLE beacons in conjunction with wearable devices to enhance the system's performance. As we discuss in Section 2, the majority of existing approaches either rely solely on wearable sensors or they use specialised infrastructure. The second contribution is the development of a data collection and labelling framework, which integrates the wearable devices and the BLE beacons and allows for the creation of labelled datasets to be used with activity recognition algorithms. Finally, the third contribution is the evaluation of our activity recognition system's classification accuracy when using different classification algorithms and feature types and the comparison of its performance to that of systems that only rely on wearable devices.

We should note that the focus of this work is to evaluate the effect of location enhancement in recognising human activities. However, instead of only providing our experimental results for the location-enhanced system, we also present results for the case where only a wrist-worn device (i.e., a smart watch) is used by the participants. This provides the baseline that can help us compare the performance of the location-enhanced system to that of systems that only use wearable devices, as discussed in Section 2.

The remaining of this paper is structured as follows. In Section 2, we discuss related literature in the area of human activity recognition using wearable devices, both commercial and custom. We continue in Section 3 with a description of our system's architecture, while Section 4 elaborates on the design of our activity recognition chain. The details of our experimental setup are presented in Section 5. In Section 6, we present our experimental results and discuss the performance of our system before we summarise our conclusions in Section 7.
