MACHINE PERCEPTION AND ROBOTICS GROUP

機械知覚&ロボティクスグループ
中部大学

Statistical learning methods for human detection require large quantities of training samples and thus suffer from high sample collection costs. Their detection performance is also liable to be lower when the training samples are collected in a different environment than the one in which the detection system must operate. In this paper we propose a generative learning method that uses the automatic generation of training samples from 3D models together with an advanced MILBoost learning algorithm. In this study, we use a three-dimensional human model to automatically generate positive samples for learning specialized to specific scenes. Negative training samples are collected by random automatic extraction from video stream, but some of these samples may be collected with incorrect labeling. When a classifier is trained by statistical learning using incorrectly labeled training samples, detection performance is impaired.Therefore, in this study an improved version of MILBoost is used to perform generative learning which is immune to the adverse effects of incorrectly labeled samples among the training samples. In evaluation, we found that a classifier trained using training samples generated from a 3D human model was capable of better detection performance than a classifier trained using training samples extracted by hand. The proposed method can also mitigate the degradation of detection performance when there are image of people mixed in with the negative sample sused for learning.