Object detection by parts using appearance, structural and shape features

This paper proposes a novel object detection algorithm with the combination of appearance, structural and shape features. We follow the paradigm of object detection by parts, which first uses detectors developed for individual parts of an object and then imposes structural constraints among the parts for the detection of the entire object. In our research, we further incorporate a priori shape information about the object parts for their detection, in order to improve the performance of the object detector. Specifically, we use an HOG-based detector for object parts whose output, together with structural constraints, is then used to seed a subsequent image segmentation step in order to delineate the potential object parts. To determine whether the segmented regions are indeed object parts, we train a part classifier using shape features of object parts and a support vector machine (SVM). The detection of the object is determined by combining the likelihoods computed with the HOG part detector, the shape-based part classifier, and the structural constraints among the parts. For validation of our object detection algorithm, we apply it to the detection of the tooth line of a mining shovel, which consists of a set of teeth with known relative position and orientation from each other, under various lighting conditions. The experimental results demonstrate that our system is able to improve the detection performance significantly when part shape information is used.