Vehicle Detection in Aerial Imaginary using a Miniature CNN Architecture

Wide area surveillance systems (WAS), which have used since years, are becoming easily accessible and widely used with the developing camera and unmanned aerial vehicle (UAV) technologies. In addition to traditional image processing methods, deep learning networks are also frequently used in object detection in wide area images. Although vehicle detection methods in WAS provide high accuracy, they also involve significant computational complexity and require hardware with intensive processing power. In this study, a low complexity and high accuracy hybrid vehicle detection method based on convolutional neural network (CNN) with support of morphological operations is presented. Results on the COWC dataset show that the proposed architecture can detect vehicles with precision of 96% even though the number of neural network parameters is just above a hundred thousands. A comparison on number of parameters with respect to commonly used neural networks reveal that the proposed architecture uses significantly lower number of parameters. Performance evaluation and the comparison of the proposed method with YOLO-v4 show the success of the presented miniature CNN network.