Neural networks modification for solving the traffic signs detection problem

The paper deals with the implementation of the traffic signs detection model on the deep neural network basis. The complexity of the problem lies in the detected objects size. The paper considers the use of modern approaches to the deep neural networks Mobile Net v2 and Tiny YOLO v3 implementation to solve the detection problem in real time. Also, a modification of the considered networks is proposed, which allows increasing the Average Precision index by more than 20%. For the networks training GPGPU NVIDIA 1080 and a signs images set called Russian traffic sign images dataset - RTSD which includes more than 15,000 frames for training and more than 3,000 frames for carrying out testing were used.