Research On The Technology Of Visual Perception Based UAV Target Recognition And Tracking

This thesis researches the problem of object recognition and tracking with the background of battlefield visual perception of Unmanned Aerial Vehicle. Based on the computer vision theory, robust and high real-time object recognition and tracking algorithm are explored aiming to complement the Deficiencies of the traditional method and improve the ability of recognition and tracking. Before these, due to the image degradation problems result from shooting conditions, image denoising and image deblurring method are investigated. Based on the need of above, we focus the following four areas:1. We propose a degraded image pre-processing method based on sparse representation. First, the theory of sparse representation is introduced and so do the image restoration model based on traditional sparse representation. Then, do some expansion to the traditional model by introducing the sparse noise model to divide the image noise from the original image. This method can improve the restoration precision effectively by suppressing the sparse noise model. Finally, according to the theory of image non-local self-similarity, we solve the problem of coefficient estimating of the sparse representation model successfully.2. We propose an object detection method based on vision saliency model. Because super-pixels can provide richer information, while the number of superpixels reduces greatly compared to pixels, so a super-pixels segmentation is adopt as preprocessing which can save computation sources in a large extent. What is more, by introducing graph model to do region fusion can overcome the interference of complicated background. At the same time, considering the object color, texture, scale and boundary are different in different images, a hierarchical model is constructed to adapt to the different cases the object in,making our method more general. At last, using guided filter to do the post-processing to the salient object detected which ensures a clearer boundary and improves the detection precise.3. We propose an object recognition method based on object proposal search. Due to the problems in traditional sliding windows, a method based on segmentation and fusion is adopted. By fuse the different regions according to their similarities, the object proposals are produced. This method can save the time which can be used to do object recognition.After that, extract descriptors of object and form a visual vocabulary by clustering. To addobject space information to the method, introduce the spatial pyramid model in which the PHOW features would be extracted. At last, we feed these object features to SVM to train an object model that can be used to recognize different categories with a high precision.4. We propose an object tracking method based on online learning model. Traditional tracking algorithm using sparse representation has a high complexity and time consuming.To enhance the tracking efficiency, we still adopt sparse representation method, but the difference is the occlusion parts and non-occlusion parts are divided in our model, which solve the part-occlusion problem successfully. In addition, PCA incremental learning is used to update the tracking templates keeping the robustness and rapidity of tracking.