As deployment of Internet of Things (IoT) devices gain momentum, there is an increased interest in implementing machine learning (ML) algorithms on IoT devices. Most of the existing ML solutions, however, rely on a central server to execute data-intensive ML models, because most devices in IoT systems do not have sufficient storage and computing resources. This paper presents a distributed Artificial Neural Networks (ANN) architecture, called PCANN, which allows execution of a complex image recognition task on a collection of resource-constrained IoT devices. Our solution separates a single ML model into multiple small modules that are executed by the distributed IoT devices. The solution effectively reduces storage and computing requirements for individual devices to store and process ML model. We design multiple PCANN models and utilize the models for human posture recognition as the case study. The experimental results show that the distributed PCANN architecture achieves comparable accuracy as the classical ANN model, while the average size of each PCANN module is largely reduced.

Digital video capturing is getting popular with the decreasing price of camcorders and the increasing availability of devices with embedded video cameras such as digital-still cameras, mobile phones and PDAs. While a raw home video is considered as visually non-appealing, having multiple recordings of the same event provides the opportunity to combine audio and video segments from different cameras for improving quality and aesthetics. Mixing content from different recordings requires precise synchronization among the recordings. In most present applications, synchronization is done manually and considered as a very tedious task. In this paper, we propose a novel automated synchronization approach based on detecting and matching audio and video features extracted from the recorded content. We assess experimentally three realizations of this approach on a common data set and make recommendations on the usability of the different realizations in practical use cases. The realizations have no limitations on the number and movement of the cameras. Moreover, they are robust against various ambient noises and audio-visual artifacts occurring during the recordings.

Combining audiovisual sequences from different cameras requires precise alignment in time. Current synchronization techniques involve using geometrical properties such as camera positions and object features. In this paper, we present a synchronization method based on detecting flashes present in the video content. Such flashes, generated by still cameras, cause a sharp bright frame in the videos. They are detected using an adaptive threshold on luminance variation across the frames. The resulting flash sequences are searched for matches. The matching flashes indicate overlapping content, and allow determining the offset time between cameras. The experimental results from flash detection and flash sequence matching show perfect synchronization in all examined cases.