TY - GEN
T1 - Convolutional neural networks of the YOLO class in computer vision systems for mobile robotic complexes
AU - Zoev, Ivan V.
AU - Beresnev, Alexey P.
AU - Markov, Nikolay G.
PY - 2019/4/1
Y1 - 2019/4/1
N2 - An important scientific direction is the development and study of computer vision systems (CVS) for mobile robotic complexes. Today, developers of CVS are most often using convolutional neural networks (CNN). For increasing the speed detection of objects on images in CVS, there has been a trend of using CNN, which are hardware-implemented on field-programmable gate array (FPGAs).This article shows that the perspective for hardware implementation on the FPGA is the tiny-YOLO CNN from the YOLO class. For reduce required FPGA computing resources in this CNN, was proposed to use Inception-ResNet modules. We was found that with high detection accuracy of objects in images with minimum resources requirements provide by the tiny-YOLO-Inception-ResNet2 network architecture. It is obtained from replacing the fifth tiny-YOLO convolutional layer of the tiny-YOLO CNN with two sequential processing Inception-ResNet modules. Also results of the study of the detection accuracy of objects using the CNN for this architecture with the lack of resource-intensive operations: batch normalization and bias from calculations were given. These studies were performed for different formats of representation numbers in the FPGA.
AB - An important scientific direction is the development and study of computer vision systems (CVS) for mobile robotic complexes. Today, developers of CVS are most often using convolutional neural networks (CNN). For increasing the speed detection of objects on images in CVS, there has been a trend of using CNN, which are hardware-implemented on field-programmable gate array (FPGAs).This article shows that the perspective for hardware implementation on the FPGA is the tiny-YOLO CNN from the YOLO class. For reduce required FPGA computing resources in this CNN, was proposed to use Inception-ResNet modules. We was found that with high detection accuracy of objects in images with minimum resources requirements provide by the tiny-YOLO-Inception-ResNet2 network architecture. It is obtained from replacing the fifth tiny-YOLO convolutional layer of the tiny-YOLO CNN with two sequential processing Inception-ResNet modules. Also results of the study of the detection accuracy of objects using the CNN for this architecture with the lack of resource-intensive operations: batch normalization and bias from calculations were given. These studies were performed for different formats of representation numbers in the FPGA.
KW - Computer vision systems
KW - Convolutional neural networks
KW - Field-programmable gate array
KW - Mobile robotic systems
KW - Object detection on images
UR - http://www.scopus.com/inward/record.url?scp=85068347615&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85068347615&partnerID=8YFLogxK
U2 - 10.1109/SIBCON.2019.8729605
DO - 10.1109/SIBCON.2019.8729605
M3 - Conference contribution
AN - SCOPUS:85068347615
T3 - 2019 International Siberian Conference on Control and Communications, SIBCON 2019 - Proceedings
BT - 2019 International Siberian Conference on Control and Communications, SIBCON 2019 - Proceedings
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2019 International Siberian Conference on Control and Communications, SIBCON 2019
Y2 - 18 April 2019 through 20 April 2019
ER -