Development of a park monitoring system for smart camera networks

Nowadays, we are surrounded by video surveillance cameras in public and private spaces. These vision systems are usually smart, i.e. with computational capabilities that allow the development of various applications. A very common application of video surveillance cameras is the Park Monitoring using Deep Learning techniques to detect the vehicles in a parking area. In particular, in this thesis work, we adopt the vehicles counting approach based on deep learning to monitor a parking lot. This approach is able to globally count the cars in the parking area without requiring any information about parking lot locations. Although deep learning is particularly effective, a factor of complexity is represented by challenging situations, due for example to the presence of shadows, variation of light and weather conditions, inter-object occlusions. To overcome these problems, it is possible to use a pair of cameras that monitor the parking lot with different perspectives and different angles of views, or multiple adjacent cameras to cover a wide area. This introduces problems related to the automatic merge of the knowledge extracted from single cameras because their fields of views partially overlap. In this thesis work, we addressed the cars counting problem and developed a solution to count cars from a video stream, using frames captured by multiple cameras. The solution combines deep learning techniques for Object Detection with a geometry-based approach, to find a homography transformation between two adjacent cameras and identify automatically the cars in their overlapping area. In particular, each camera uses the Region Based Convolutional Neural Network Mask R-CNN to detect cars in its park portion. To test the solution, we used the existing CNRPark-EXT dataset, composed of images taken by nine smart cameras located in the campus of the National Research Council (CNR) in Pisa and covering challenging scenes. We propose two different approaches to implement the solution: a centralized solution, in which the system is composed of some cameras that send the captured frames to a central server, which sequentially performs cars detection on them and merging those results, or a decentralized solution, in which the system is composed of some smart cameras that communicate with each other and perform the detection and merge tasks on-board. We finally show the results obtained and how a real multi-camera car counting system could be implemented