The article consider the influence of illumination and distance on the recognition quality for various models of neural networks of embedded systems. The platforms on which the testing was carried out, as well as the models used, are described. The results of the study of the influence of illumination on the quality of recognition are presented.
Keywords: artificial intelligence, computer vision, embedded systems, pattern recognition, YOLO, Inception, Peoplenet, ESP 32, Sipeed, Jetson, Nvidia, Max
The article discusses methods for optimizing floating point calculations on microcontroller devices. Hardware and software methods for accelerating calculations are considered. Algorithms of Karatsuba and Schönhage-Strassen for the multiplication operation are given. A method for replacing floating-point calculations with integer calculations is proposed. Describes how to use fixed point instead of floating point. The option of using hash memory and code optimization is considered. The results of measuring calculations on the AVR microcontroller are presented.
Keywords: floating point calculations, fixed point calculations, microcontroller, AVR, ARM
The article discusses the application of machine vision methods for embedded systems using modern microcontrollers. Machine learning methods that are used in embedded systems to solve recognition problems, as well as neural network models, are described. The use of trained models for solving image recognition problems in embedded systems is proposed. The architectures of YOLOv3 and R-CN neural networks are compared. The Jetson TX2 hardware platform is considered. The results of comparing the calculation speed for different modes of the device are presented.
Keywords: machine vision, neural networks, artificial intelligence, embedded systems, pattern recognition, YOLO, RCN, Jetson, Tensorflow