Research On Real-time Object Detection Using Adaptive Multi-window

Object Proposal is a window sampling means in digital image processing, in order to overlap all regions of interest using windows as little as possible, it assumes there are common features between the regions of interest, first produce windows according to a certain strategy, then extract common features of object to design a objectness estimation method, finally sort and filter these windows. Comparing to sliding window algorithm commonly using in image processing, object proposal can produce windows with less quantity and higher precision, has become an effective approach to increase the efficiency in recent object detection methods. Besides, object proposal can also optimize the sampling procedure in other image processing like motion detection and object tracking, as image processing becoming more and more complex and the increase in the demand for computational efficiency and precision, research on object proposal has important significance.This thesis analysis and compare the existing object proposal methods, then settle for a sampling idea which filter first and optimize second aim at the low area overlap ratio and long computation time of object proposal. First design a quantization search strategy based on overlap ratio to produce windows, select gradient matrix as feature, train a liner two classification model using cascade SVM; then analysis the relationship between edge and object, design an object estimation method based on edge which is used to sort and optimize the generated windows. In order to prove the efficiency of this algorithm, the sampling results in VOC2007 dataset is input into object detection method to do object detection task.Compared with other object proposal methods, this algorithm hold high calculation speed and detection recall, also solved the problem of low accuracy in calibration windows. Through the comparison and analysis of various performance metrics on the VOC2007 dataset, this algorithm had fastest calculation speed next to BING(0.1s) and a good detection recall in the high overlap ratio over 0.7, it also improved the mean average precision for object detection method.This thesis not only demonstrated the search strategy based on overlap ratio, but also designed a high reliability objectness estimation method based on edge, which proved that the edge which intersected in window will interfere with the calculation of the physical properties. It also proved that the high precision and low number of sample windows can effectively improve the accuracy of the object detection algorithm. In other Image processing algorithm using sliding window method as means for searching and matching, this self-adaptive multi-window algorithm can be used as a alternative sampling method, reducing the number of matches and exclude the interference of negative samples.