Developing a real-time human and vehicle detection system for monochrome wide area surveillance images

Wide-area aerial images are one of the main sources used for reconnaissance and surveillance. Target detection from very large-sized aerial images is a remote sensing application that requires high processing power. In order to have a reasonable wide-area surveillance system, computational complexity on image processing should be minimized as much as possible. On the other hand, in recent studies, deep learning based approaches are preferred instead of traditional methods with relatively low complexity. The use of a deep learning based target detection system in wide-area images reveals the need for miniature neural network architecture. However, most of the studies on target detection from aerial images use convolutional neural networks with high computational complexity. In this thesis, a miniature convolutional neural network was designed for view-based target detection, and high accuracy vehicle detection was performed on aerial images with the presented system approach. The proposed system has been tested on various datasets for human and vehicle detection, and it has also been evaluated with a compound dataset, which is a collection of these used datasets. The results show that with a fully convolutional deep neural network with a very low number of parameters, the performance of the state-of-the-art model can be achieved and even better results can be obtained. The low number of parameters in the proposed method allows the system to be run on parallel programming hardware such as FPGA or to fit into ASIC designs.