Vehicle Detection Algorithm Based On Visual Saliency On Aerial Images

Target detection is a fundamental and important research theme in Computer Vision.Target detection has been applied to many fileds such as target recognition,video surveillance,retrieval based on image content an,human computer interaction and so on.With the deepening of psychological research,the researchers find that the human visual system can analyze the information of outside world efficiently and locate and react to the areas of interest in priority.This visual cognitive mechanisms bring new methods and ideas to target detection.Based on characteristics saliency analysis, and combined with a classification, The paper implements a aerial vehicle detection algorithm.The paper mainly completed the following work:First, as a resolution to the vehicle detection on aerial images, according to the characteristics of the human visual system, a vehicle detection algorithm based on characteristics saliency analysis has been proposed. To obtain a feature response map, the bottom edge of the vehicle information is used to do image convolution on the input image. Then calculate the spectral residual of feature response maps on different scales to obtain saliency response maps. Finally, in order to obtain the final saliency map, multi-scales saliency maps has been merged into one map and temporal enhancement and spatial enhancement method has been used. By thresholding processing and K-means clustering algorithm to obtain the target candidate region.Second,saliency detection algorithm has a problem that there are lots of false object detection.In order to solve this problem, we use a multilevel HOG feature and cascade adaboost classifier to detect vehicle on aerial images.we constructe a cascade classifier to solve the problem of uneven samples.In order to solve the case of multi-scale vehicle detection,we also present an image pyramid approach to search vehicle in multilevels.In summary.the vehicle detection system consists of a saliency detection module and a classfier detection module.The whole system performs well in real aerial scene.