Object Detection Based On Hough Transfrom

The Hough transform is popular studied in computer vision techniques. It wasinitially suggested as a method for parametric geometrical shape detection, such as lineand circle in images. The generalized Hough transform was developed by Ballard to detectarbitrary shapes. In recent years, the methods which combined Hough transform andclassifier were successfully adapted to the problem of category-level object detectionwhere they have obtained state-of-the-art results for some popular datasets. However, dueto occlusion and clutter backgrounds, the detection of the single or multiple targets rapidlyand accurately is still tough. Based on the analysis of the related domestic andinternational research results, this paper studies the object detection algorithm based onHough transform.Firstly, to detect circular target in work-piece image accurately, we makecorresponding improvement for the existing random Hough transform. The noise isfiltered and the image arc edge is extracted by8neighborhood contour tracking.According to the prior knowledge, the elimination of false targets based on the range ofradius would reduce calculation. The coarse position of center is located rapidly byapplying the geometrical features of circle. Then, use the generalized least squares withinthe scope of the ring to extract the circular parameters accurately in sub-pixel level basedon certain criterion. Namely, adopt the coarse-to-fine detection for the location of circulartarget.Secondly, combining Hough transform and Random forest for object detection, thedataset of training images contains HOG feature information, class information and spatialdistribution information. The Hough forest is builded under the supervision of categoryinformation and all leaf nodes form a tree structure discriminative codebooks model. Then,the leaves that the testing image patches arrive in are then used to cast probabilistic votesabout the possible location of object. The appearance of objects of the same class innatural images varies greatly due to intra-class differences, illuminations and backgroundas well as object articulations. So a structure model of Hough forest is constructed bysignificant object parts manually selected and spatial distribution information relative to object centre.Finally, the detection of multi-object and occluded object is accomplished bycombining Hough transform and probabilistic model. We develop a new probabilisticframework for object detection which is related to the Hough transform to explain therelationship between the voting elements and hypothesizes. In addition, in the objectsdetection of Hough image, a greedy algorithm is used for iterate Hough vote instead ofnon-maximum suppression heuristics. As a result, the method bypasses multiple peakidentification and has high accuracy.