**1. Introduction**

During recent years, the interest in indoor human location has increased due to the large number of applications in various fields, such as security surveillance, activity monitoring, behavioral analysis, and healthcare [1].

Traditionally, when it is necessary to locate people indoors, radio-based technologies are used, which can be affected by the characteristics of the environment and also force target users to carry specific devices. An alternative to these technologies is video-based localization using security cameras, which are increasingly common in buildings and public places. Due to advances in computer vision and Deep Learning (DL), the detection of people on video is more reliable.

The localization and tracking of people is usually performed in two steps: people are first detected in each individual frame to obtain their position in the image. Then, these detections are associated across frames to obtain the path followed by each person.

Typically, these tasks are performed by processing the video from each camera in a centralized way. However, it is possible to perform this processing in a distributed manner due to the advances in edge-computing. Cameras can use AI accelerator chips that allow for fast and low-power neural network inference. Several chips are available on the market, such as Google Coral, Intel Movidius, or Nvidia Jetson.

In this work, we focus on the task of detecting people in security camera images by performing the processing on an embedded device with a Google Coral's Edge TPU. The tracking methods that can be applied to the obtained results are not addressed.
