Parking Spot Occupancy Classification Using Deep Learning

In today’s world, where a car is present in almost every family, the parking problem plays an extremely important role. Parking is one of the most important factors in modern transport infrastructure, because it allows to save the time of both drivers and passengers, to increase the level of comfort and safety of road trips. In Ukraine, this problem is especially relevant, since nowadays it is going through the process of improving its parking infrastructure.The paper examines the problem of parking in large cities, proposes a system for recognizing occupancy of parking spots using computer vision. Such system would use camera feed to track the occupancy of each parking space within a slot. Its benefits would include ease of scalability, saving time of drivers and passengers, automation of parking payment and detection of unpaid parkings. In addition, it makes it possible to easily collect statistics about the busyness of various areas throughout the day or week.The paper also describes the algorithm of classifying the parking spot, as well as a possible architecture that the system may have.Possible problems in training a computer vision model for building the proposed system are considered. Firstly, the available parking datasets are lacking images collected in snow conditions or during nighttime. The hypothesized solution is to use vehicle detection datasets, the number of which that are publicly available is considerably bigger. Another problem is that classification accuracy drops drastically when using different images in train and test dataset. The hypothesized solution here is to apply incremental learning to improve the model as it is being used in a real-life scenario.