Performance of YOLO based traffic analysis in high resolution video stream

With the development of advanced deep learning techniques, object detection and classification in digital images has entered a level once reserved only for humans. Analyzing the data in a video stream has always been a challenging process, but the continued development and miniaturization of electronics makes it possible to perform this task based on popular hardware platforms. Today, complex detection mechanisms are available through pre-trained open solutions such as YOLO. For this reason, this library is widely used in many applications. The paper focuses on analyzing the performance of recognizing and counting vehicles in road traffic, using the YOLO library (version 5 and 8). The study analyzed the performance of object detection in a video stream for different atmospheric conditions using the CPU and GPU. In addition, the efficiency of counting classified vehicles in practical application was verified. The results obtained for good quality video (FullHD/UltraHD) indicate that without the use of modern GPUs, analysis and detection of vehicles without skipping frames is not yet possible for more complex problems. The use of popular CPUs, even with more than a dozen cores, made it possible to achieve processing speeds on the order of 2 fps in case of the ”x” YOLO version and up to 20 fps for the ”s” version. Real-time analysis of a higher-resolution video stream is obviously possible using frame skipping, reducing the analyzed area and applying less complex preprocessing (filtering) techniques.