Object detection and recognition method based on evidence accumulation and gabor features

Abstract

In this paper, we propose an invariant object detection and recognition method based on evidence accumulation and the Gabor transforms features. Gabor transform is an image transformation, which performs local frequency analysis. Gabor transform coefficients represent localized shape or textual information such that they can be used to encode and detect object part features. In previous works Gabor transform features are used as an object representation to find alignments between image and memorized object model. The application domain of Gabor features in previous works was confined to cases where objects and object models are roughly aligned. Evidence accumulation detects global object pose invariantly by detecting local parts and integrating the partial information using geometric relationships between parts. Such approaches include Hough transform, generalized Hough transform, pose clustering, and evidence accumulation. Previous evidence accumulation approaches explores only simple local features such as edge, corner, and contour segments. In contrast to conventional evidence accumulation researches, the proposed method uses Gabor transform features to detect and match object parts. After the global object pose has been determined, matching between object models and the image is evaluated by Bayesian method. The experimental results show that our algorithm can robustly detect objects in cluttered environments with invariance to translation, rotation, scaling, occlusion and small deformation. We also show the experimental performance of object recognition scheme using object models based on local shape encoding by Gabor features.