The detection, classification and counting of vehicles through images of video cameras is an actual task for: analysis of traffic congestion for traffic management optimization; detection of violations of the weight regime; analysis of the road deteriorations or objects of transport infrastructure . High complexity of automated identification is one of the problems of transport classification in streaming mode [1]. The use of manual classification is routine and costly from human factor standpoint. Automated classification requires the implementation of machine learning algorithms and the formation of training samples. Algorithm analysis for recognizing transport types showed the neural network approach could be advanced [2], and convolutional neural networks (CNN) and groups of several deep neural networks were presented as the most effective tools [3–6]. Convolutional neural network, or convolution network, is a multi-layer perceptron contrived for recognizing two-dimensional surfaces with a high degree of invariance to transformations, scaling, distortions and other kinds of information [7]. The use of classical neural network architectures (Hopfield models, Kohonen’s selforganizing maps, Elman’s recursive networks) for the problem-solving in the conditions of the video stream is not effective because of external factors affect sensitivity (changing the camera angle, scale, speed of movement of cars). A convolutional neural network doesn’t have this disadvantage, so its implementation would be effective for a vehicle class recognition system.