Deep Learning in Vehicle Pose Recognition on Two-Dimensional Images

Abstract

The paper describes usage of deep neural network architectures such as VGG, ResNet and InceptionV3 for the classification of small images. Each image may contain one of four vehicle pose categories or background. An iterative procedure for training a neural network is proposed, which allows us to quickly tune the network using wrongly classified images on test sample. A dataset of more than 23,000 marked images was prepared, of which 70% of images were used as a training sample, 30% as a test sample. On the test sample, the trained deep convolutional neural networks are ensured the recognition accuracy for all classes of at least 93.9%, the classification precision for different vehicle poses and background was from 85.29% to 100.0%, the recall was from 81.9% to 100.0%. The computing experiment was carried out on a graphics processor using NVIDIA CUDA technology. It showed that the average processing time of one image varies from 3.5 ms to 15.9 ms for different architectures. Obtained results can be used in software for image recognition of road conditions for unmanned vehicles and driver assistance systems.