IEEE Second International Conference on Data Stream Mining & Processing August 21-25, 2018, Lviv, Ukraine 978-1-5386-2874-4/18/$31.00 ©2018 IEEE 534 Embedded Vision Modules for Text Recognition and Fiducial Markers Tracking Ievgen Gorovyi It-Jim Kharkiv, Ukraine ceo@it-jim.com Valerii Zozulia It-Jim Kharkiv, Ukraine ceo@it-jim.com Vitalii Vovk It-Jim Kharkiv, Ukraine ceo@it-jim.com Maksim Shevchenko It-Jim Kharkiv, Ukraine ceo@it-jim.com Dmytro Sharapov It-Jim Kharkiv, Ukraine ceo@it-jim.com Abstract—In the paper, two examples of embedded vision modules are described. Firstly, it is demonstrated how fiducial marker tracking algorithm can be adopted for operation on Raspberry Pi. Usage of proposed ideas allows to achieve around 60fps speed of binary marker tracking. Secondly, we describe the problem of text detection and recognition in outdoor environment. Experimental results indicate on acceptable results and good potential to provide low-cost and efficient embedded vision system for this purpose. Technical details of both embedded vision modules are comprehensively discussed. Keywords—computer vision, Raspberry Pi, fiducial markers, tracking, text recognition. I. INTRODUCTION Computer vision (CV) is a rapidly growing discipline making machines to percept and understand their surroundings as humans do [1]. There are a lot of practical applications of CV in medicine, industry, entertainment and many more [2]. CV algorithms can be run on different hardware: desktops, mobile phones, various digital signal processing (DSP) units. A particular interest is related with usage of low-power hardware such as Raspberry Pi [3]. Indeed, Raspberry is light weight, cheap and widely available in the market. Embedding of CV solutions transforms it into mobile autonomous intellectual system. Fig. 1 contains an example of Raspberry (Fig. 1a) and its setup with camera (Fig. 1b). (a) (b) Fig. 1. Raspberry PI 3 and camera. (a) Raspberry PI Model 3B, (b) Raspberry with camera Research community made a lot of CV experiments with Raspberry Pi. Example of multiple objects tacking can be found in [4]. In [5] a compact stereo-vision system is described. A full stereo matching pipeline is constructed allowing to use the system as a depth-meter for outdoor scenarios. A specific use case in given in [6], when bees behavior is analyzed using embedded vision. Other examples include face recognition [7], license plates detection and recognition [8], autonomous cars applications [9], robotic assistants [10] and many more. In the paper, we study two important problems. Firstly, it is demonstrated how to integrate and optimize fiducial marker recognition algorithm for Raspberry. Secondly, we describe the initial experimental results on number plate recognition as a separate embedded vision module. Section II contains description of algorithm for fiducial marker recognition. Developed optimization steps are described as well. Section III contains information about text detection and recognition on Raspberry. Experimental examples are included in each section. II. FIDUCIAL MARKERS RECOGNITION In this section, we describe a pipeline for recognition of fiducial markers. Firstly, common steps for such process are explained. After that, some specifically developed improvements are discussed. A. Main Steps of Marker Recognition Algorithm Binary markers are used in many applications including robotic navigation [11]-[13], augmented reality [14] and logistics [15]-[17]. Examples of typically used binary markers are shown in Fig. 2. Fig. 2. Example of binary markers Lviv Polytechnic National University Institutional Repository http://ena.lp.edu.ua