Desert Vehicle Detection Based On Visual Attention And Pulse Coupled Neural Network

Object detection and recognition in remote sensing images is a hot topic in research field recent years, especially vehicle detection, which has great significance in transportation management and vehicle rescue. In vehicle rescue field, because of the high temperature, strong sand storms and weak GPS signal in the complex and variable desert environment, vehicles are more prone to break down. However, traditional vehicle detection algorithms could not satisfy desert vehicle detection requirement. Firstly, the resolution of remote sensing image is very low that could not satisfy resolution requirements of many methods. Secondly, vehicle unit is small in remote sensing image so that it is hard to extract vehicle model or other information which is important to traditional detection algorithms. Thirdly, interfering factors of background is similar to vehicles, such as plants and hills, causing regions of interest (ROI) hard to locate. For these reasons, traditional vehicle detection algorithms have disadvantage of low recognition rate and high false alarm rate in desert sense, which cannot meet the requirements of high computational efficiency and accuracy of desert vehicle rescue. In this paper, we propose quaternion visual attention model with adaptive background channel to identify desert vehicle area with pulse coupled neural network. After that, scale-invariant feature transform and hierarchical discriminant regression are used to further lower the false alarm rate of detection. The main innovations in this paper could be represented as follows:1. We propose a desert vehicle identification method based on visual attention model. We propose a quaternion visual attention model with adaptive background channel to calculate saliency map based on image phase spectrum. Saliency map present saliency area of image by different intensity could smaller search area of vehicle which improve detection speed.2. We propose a desert vehicle detection method. Aiming to lower false alarm rate of desert vehicle detection. We further introduce pulse coupled neural networks to locate particular area of interests (ROIs) of vehicle. After that, extract features of ROIs with scale-invariant feature transform (SIFT) and identify vehicle areas by hierarchical discriminant regression (HDR) tree. This procedure would improve false alarm rate and detection rate significantly.3. We propose two visual attention model with learning ability. Due to limitation of quaternion visual attention model is just a bottom-up (data-based) model. We use least square method and support vector machine to introduce learning ability and pre-knowledge information to original visual attention model. The improved model could calculate object-based saliency map which further improve detection rate and false alarm rate of desert vehicle detection. Further, proposed model could be applied to sea object detection situation.