The system for automatic recognition and classification of cars in traffic is a device whose task is to independently in real-time from the available video signal to detect the car, recognize the license plate, determine the color and model and make the car. The paper describes in detail the principle of operation of convolutional neural networks which are the basis for the realization of the practical part of the work. Traditional methods such as the Viola-Jones algorithm are described, followed by modern methods of object detection. Single-stage detectors based on convolutional neural networks were used. Cars are detected using the YOLOv4 detector. The output from YOLO is then forwarded for further classification. The color is determined using the KNN algorithm, and the model and make of the car using an SSD detector. The last is the detection of the license plate using the OpenCV method, so the characters from the table are read using Tesseract. Precisely because of the simplicity and speed of operation of these subsystems, they were chosen to work on a Raspberry Pi embedded computer. Car information is printed on the monitor. The entire system was tested on images and video. Car detection accuracy is 88,5% and model classification is 78,5%. Color accuracy is 71,5% and 51,5% is licence plate accuracy. The system operates at 1 image per second (1fps). The disadvantages and advantages of the whole system are described and possible improvements of the system for further development are given.