Object Detection Using Hough Transform And Conditional Random Field Model

Object detection is an active research area in such fields as pattern recognition andcomputer vision. Due to occlusion, different viewpoints, and variations of illumination andscale in cluttered real-world scenes, the objects of the same class usually exhibit great visualvariation, which makes the task challenging. Object detection from an image and video isusually performed using the sliding-window-based approaches. Recently, Hough transform-based approach demonstrates an alternative and interesting way for the task of objectdetection. It was originally used to detect lines, and its generalized form was developed todetect predefined shapes.Implicit Shape Model extend the underlying idea behind Hough transform andemployed the so-called Hough voting of local features in3D Hough hypothesis space. Bysimply summing all the votes, it searches locate peaks in the Hough space to detect objects.Despite the good performance of Implicit Shape Model, it also suffers from some drawbacks:it assumes every Hough votes of each feature is independent. From a probabilistic point ofview, this independence assumption is quite rude, since the neighboring features of an imageare actually related.Conditional Random Filed is a type of discriminate probabilistic model, it inherits theadvantages of Hidden Markov Model, and meanwhile removes the strong independenceassumption of Hidden Markov Model. Inspired by Conditional Random Field, in this paperwe use it to model the relationship between the voting features and the hypotheses in theHough transform so that the conditional probability of hypothesis is not only depend on theset of voting features, but also the other hypotheses in neighbors. We formulate theprobabilistic model by random forests classifier and parzen-window estimate. The mainprocedures as following: 1）Hough voting based on random forests. We build the random forests classifier usingthe SIFT features from the positive and negative examples. For each leaf node, westore the offsets and class information of these features which reach this node, thusthe set of leaf nodes can be viewed as a discriminate codebook. We then use thecodebook and generalized Hough transform to estimate the object’s location afterthe SIFT features from test image are classified by random forests.2）Parzen-window estimate. We use the kernel function to formulate the dependencybetween hypotheses, which performed by Scale-Adaptive Mean-Shift method.3）Greedy algorithm based on maximum-a-posteriori inference. In this paper, we usethe MAP inference to formulate the probabilistic model. We iteratively update theHough voting space partially after finding the maximum score to detect object untiltermination.We evaluate the performance of our approach with three public test sets (UIUC-Cars,TUD Motorbikes, Weizmann Horse) and a domestic airport test set we collected ourselves.The experiment result shows that our method achieves better performance compared withImplicit Shape Model, even when objects are occlusive.