Category: Research

In the paper, we analyze the problem of automatic room floor segmentation. For this purpose, we consider several classic computer vision algorithms as well as some of the deep convolutional neural network architectures. The segmentation results are illustrated and compared. An idea for combining two groups of methods is proposed. It is demonstrated that a proper fusion provides the best segmentation quality.

The paper presents basic concepts of augmented reality applications and challenges in building them in the web. We describe the technical and algorithmic stack required to develop, implement and deploy the augmented reality application. Theoretical concepts behind marker detection and tracking are discussed. Two different pipelines are implemented: server-based with algorithms execution in the cloud and completely front-end solution that runs on a user device. We show advantages and disadvantages of each approach and analyze experimental results as well.

In the paper, we describe the technical details of a multi-player sports tracker system. We demonstrate that proper in-depth analysis of video frames sequence may provide a lot of useful information required for sports analytics. Object detection and tracking steps are analyzed. Novel ideas for efficient filtering of false detections and irrelevant tracks are proposed.  Also, we show an example of how the object tracking information used for the regions of interest location allows streaming sports events without human operators. At last, important practical implementation details, as well as initial experimental results, are discussed.

3D mapping techniques have a large variety of applications from entertainment to military and medical fields. However, obtaining a well-refined 3D model from a set of images without the usage of depth sensors is a big challenge. In the paper, we analyze the main components of the 3D reconstruction pipeline allowing us to get detailed models of outdoor objects from drones. In particular, we experiment with algorithms required for structure from motion and point cloud densification. We demonstrate that proper local feature extraction, matching and verification directly affect a final model quality. Analysis of two existing 3D reconstruction frameworks (MVE and COLMAP) is conducted. Initial experimental results are shown.

Indoor positioning systems in GPS-denied environments are rapidly becoming popular. Various options are commonly available (BLE, Wi-Fi, ultra-wideband, ultrasonic, etc.). The key challenge is to provide accurate, and stable real-time user location at a low cost. In this paper, we present the research and production details of the developed hybrid indoor localization and navigation system (HILN). The proposed technical solutions are based on cheap Bluetooth beacons and mobile sensors. In particular, we describe two separate positioning pipelines for open spaces and narrow environments. The scheme of efficient fusion of inertial navigation system (INS) and BLE navigation system is proposed. All the developed solutions are integrated into the mobile indoor software development kit (SDK). Its main components are briefly mentioned. Our mobile positioning system provides 1-2m accuracy and works on Android and iOS devices on a real-time basis.

The modified You only look once (YOLO) network architecture that allows one-dimensional direction estimation along with classic object detection in real time, is considered in the task of street traffic surveillance from unmanned aerial vehicles. The key feature is a modified output fully connected layer with additional orientational parameters. It has been shown that this network can estimate the direction of vehicles on a custom testing dataset with photos.

AR technology is rapidly growing direction. Nevertheless, mobile and embedded AR still faces a lot of real-life problems and challenges. The talk will be devoted to review of AR principles and applications. The key part will contain the analysis of available technical solutions for mobile AR. Indeed, quite a lot of software engineers use the 3rd part frameworks (ARkit, ARcore, Vuforia, etc.) without putting enough attention into algorithmic and implementation details. In the presentation, we will consider the state-of-the-art algorithms for marker detection and recognition. Moreover, several fresh methods for real-time marker tracking will be discussed. In addition, such hot topics as markerless tracking, sensor fusion and SLAM will be covered as well. Finally, several real demos from custom mobile AR SDK will be shown.

It is known, that nowadays almost every indoor positioning and navigation system (IPNS) consists of a radio signals part (Wi-Fi or BLE) and a part based on smartphone inertial sensors. Both parts contain a number of challenges complicating a precise user positioning using mobile phones or tablets. In the paper, we describe several contributions. Firstly, a problem of BLE packets recovering is considered. A specific version of a Kalman filter for received signal strength indicator (RSSI) data analysis is developed. The proposed modification allows recovering lost data as well as providing sufficient signal smoothing. Secondly, a custom step detection procedure based on an inertial navigation system (INS) is developed. Unlike to a common solution based on the thresholding of linear acceleration amplitude, an advanced version of the detector is highlighted. Finally, a hybrid indoor localization and navigation (HILN) system developed on the basis of a particle filter (PF) and the proposed modifications for BLE and INS parts is described. Experimental results are provided.