In this paper, we present an object detection technique that uses scale invariant local edgel structures and their properties to locate multiple object categories within a range image in the presence of partial occlusion, cluttered background, and significant scale changes. The fragmented local edgels (
key-edgel,
ek) are efficiently extracted from a 3D edge map by separating them at their corner points. The 3D edge maps are reliably constructed by combining both boundary and fold edges of 3D range images. Each key-edgel is described using our scale invariant descriptors that encode local geometric configuration by joining the edgel to adjacent edgels at its start and end points. Using key-edgels and their descriptors, our model generates promising hypothetical locations in the image. These hypotheses are then verified using more discriminative features. The discriminative feature consists of a bag-of-words histogram constructed by key-edgels and their descriptors, and a pyramid histogram of orientation gradients. To find the similarities between different feature types in a discriminative stage, we use an exponential χ
2 merging kernel function. Our merging kernel outperforms the conventional
rbf kernel of the SVM classifier. The approach is evaluated based on ten diverse object categories in a real-world environment.
抄録全体を表示