Automatic Visual Classification of Parking Lot Spaces: A Comparison Between BoF and CNN Approaches

Abstract

Computer vision has a wide and diverse range of applications nowadays. A particular one is automatic detection of parking lot occupancy, where a computer has to identify whether a parking lot space is empty or occupied. As in any visual classification problem, detecting parking lot spaces relies on the existence of a representative visual dataset. This problem of binary classification is commonly approached using features with adequate level of invariance to changes in illumination or rotation, that allow feeding these features into classifiers such as the SVM. Most used approaches are based on the use of convolutional neural networks, some times based on pre-trained models which in general have quite high performance. however several of these methods are tested with common experiments that do not take into account the variations that occur when training with different combinations of angles, lighting variations, and weather types. That is why in this paper we present a comparison between two approaches to solve the problem of parking lot classification with two methods: Convolutional Neural Networks and Bag of Features. In this paper we show how to use the standard Bag-of-features model to learn a visual dictionary, and use it to classify empty and occupied spaces. Results are compared with CNN approaches, emphasizing on accuracy, sensitivity analysis, and execution time.