Research On Low-altitude Airborne Vehicles Detection Based On Visual Attention And Classification

With the rapid development of our urbanization, the burden of the road transportation continuously improved and the traditional manual surveillance method cannot meet the requirement of developing. Therefore, more attention has been paid to the vision-based intelligent transportation. Especially, the low altitude airborne vehicle detection system, which takes the advantages of wider view and higher mobility, has become one of the hottest topics. It aims at detecting the vehicle using the camera fixed on unmanned aircraft for relieving the traffic jam and incident collisions. Obviously, the vehicle detection system is a monitoring system with moving camera. Compared with other monitoring systems with static camera, the airborne vehicle detection system has the following difficulties: scene complexity, platform movement and high requirement of detection performance especially the high detection speed. Compared with detection system of high altitude, unprecedented difficulties are brought because of changeable perspective.For low-altitude airborne vehicle detection of urban traffic, the direct application of the traditional image processing techniques to the problem may result in low detection rate and cannot meet the requirements of real-time applications. Moreover, the existing single, cascade and parallel classifiers can hardly satisfy the comprehensive requirements of the detection system in detection rate, false positive rate and detection speed, and cannot handle the difficulties in vehicle detection with different orientations.Considering the disadvantage of existing airborne vehicle detection method with single optical camera, a new and efficient framework, which based on visual attention and classification, is proposed in this paper according to Bayes's model. This framework is a coarse-to-fine process which splits the features into two groups. The first feature set imitates the visual attention of primate and take the advantage of its processing speed. This stage of attention focus extraction consists of two alternating processes, attention shifting and extent tracing, which is used to locate potential vehicle regions in a short time. Therefore, mass computational consumption of the potential vehicle region extraction can be averted in this stage. In addition, the cascade classifier trained by Adaboost algorithm, which is sensitive to vehicle orientation, is proposed in this paper to extract the most possible orientation. Moreover, an efficient and well trained tree-like classifier is used to refine the candidate regions. Our framework takes the advantage of processing speed of visual attention and accuracy of classifier while averting each disadvantage. This method, which has fast detection speed, low false positive rate and high detection rate, can be used for handling the difficulties in vehicle detection with different orientations. Compared with other representative algorithms, the real scene based experiment proved the advantage and effectiveness of our method.