Research On Key Technology Of Visual Object Detection And Segmentation

Object detection which estimates location and scale of a given object category is an important problem in computer vision. The research on object detection is multi-disciplinary, so it has great scientific research value. In addition, the object detection as an essential component of computer vision system also has great practical application value. In the past two decades, object detection technology has developed rapidly and obtained many valuable research results. Window search object detect method which has simple framework and high detection precision receives great interest from researchers. However, window search object detect method still has some drawbacks, including low detection speed, high training complexity, repetition detection for single object, and low positioning accuracy. To solve the weaknesses, this dissertation carried out thorough research and proposed some related solutions.At the beginning of this dissertation, we elaborated the background and significance of object detection research. Then, the basic framework and related key technologies of window search object detection method were introduced. On the basis of existing methods, this dissertation focused unsolved problem of window search object detection method and covered the following works:1. Propose a new window search method based on region growing.Window search method usually generates a large number of candidate object windows to prevent from missing object, which lead to a heavy computational burden and limit detection speed. To solve the problem, we proposed a new window search method based on region growing. This method generates candidate object window using segmentation cues and prior knowledge. This method which generates a small number of candidate windows can effectively improve detection speed.2. Propose a fast object detection method based on cascade selective window.To further improve detection speed, this paper proposed a fast object detection method based on cascade selective window. This method increases detection speed in terms of three aspects: 1) The method compresses high-dimensional image channel feature by a sparse projection matrix which reduces the time of feature extraction; 2) The method designs a soft cascade SVM classifier; 3) The method introduces selective window search strategy into cascade classifier to further reduce the amount of computation. The experimental results show that the proposed method outperforms state of the art methods and effectively speed up object detection. Furthermore, another significant advantage of this method is that it allows for faster training.3. Propose a new object detection method based on local region sparse representation.In order to reduce complexity of object model training, this paper proposed a new object detection method based on local region sparse representation. Firstly, the proposed method extracts many local regions of object as training samples. Secondly, a discriminative dictionary whose atoms have explicit relations with local regions is learned. Thirdly, the method determines each candidate window whether a particular local region appears using the response of its sparse coding. Finally, object location is obtained using position constraints and detection results of local regions. Experimental results show that the proposed method which takes less training time and only needs a small number of positive samples to learn model effectively reduces the training complexity. Meanwhile, this method draws on the idea of deformable part based model, so it is robust to occlusion, deformation of object, variability of perspective.4. Propose a new detection windows fusion method based on theory of heat diffusion.Window search object detection method would cause repetition detection for single object. To solve the problem, we proposed a new detection windows fusion method based on theory of heat diffusion. The method treats each preliminary window as a location in system. Then heat conductivity between two locations is calculated by detection scores and overlapping area of corresponding windows. Finally, the detection windows fusion task is modeled by temperature maximization on linear anisotropic heat diffusion, of which the temperature maximization with finite K heat sources corresponds to K final windows. Experimental results show that this method not only deletes overlapping detections, but also rejects false positives and prevents interference between adjacent objects.5. Propose a new object detection method based on multiple cues.Window search object detection method can’t obtain exactly object segmentation results. To solve the problem, this paper proposed an object detection method based on multiple cues. This method combines results from detector, pairwise relationships between superpixels, object probability and homogeneity probability of segments into an objective function. The final object segmentation result is obtained by minimizing the objective function. Experimental results show that our approach can robustly handle the case of complex background, multiple object, and great variability in size.At the end of this dissertation, we summarize the strengths and weaknesses of the proposed methods, and then introduce the subsequent work.