The article deals with the system of parking space analysis, based on the processing of video monitoring through image segmentation, performed with the help of an artificial neural network (ANN). Two key algorithms included into the parking space analysis system are described in detail: 1) the algorithm of the image quality analysis; 2) the algorithm of the adaptive image analysis. The main purpose of the development and implementation of the first algorithm is to evaluate the feasibility of ANN image analysis, based on the recognition of object boundaries. The idea of the second algorithm-adaptive image analysis is as follows: to analyze images of different size and approximation (scale), by bringing the input image to a single dimensional format, followed by the approach of splitting into patch. After ANN analyzing the resulting parts of the image, the algorithm again combines them into a whole picture, identical to the input format. The final stage of the system proposed by the authors is an algorithm for the classification of multiple parking spaces, based on the combined analysis of the segmentation mask obtained from the ANN and a pre-prepared polygonal grid of parking spaces in the image. The paper describes the data used to achieve the obtained results and the characteristics of the training of the semantic segmentation model, giving the statistics of the performance.
Keywords:
computer vision, video monitoring, image analysis, segmentation, machine learning, data analysis
