局所特徴量の関連性に着目したJoint特徴による物体検出(オーガナイズドセッション1)  [in Japanese] Object Detection by Joint Feature Based on Relations of Local Features  [in Japanese]

本稿では,複数のHOG特徴量間の共起を表現するJoint特徴を用いた動画像からの物体検出法について述べる.Joint特徴は,組み合わされた2つのセル間のHOG特徴量の共起を表現し,1段階目のReal AdaBoostにより組み合わせる.次に,生成されたJoint特徴候補のプールを入力とした2段階目のReal AdaBoostによって最終識別器を構築する.これにより,単一のHOG特徴量のみでは捉えることができない物体の対称的な形状や連続的なエッジを捉えることが可能となる.Joint特徴の有効性を示すために,人と車両を検出対象として評価実験を行い,Joint特徴の有効性を述べる.さらに,異なる解像度のHOG特徴量間の共起や時空間特徴量の共起,TOFカメラから得られるデプス情報との共起の効果についても述べる.

This paper presents a method for detecting objects in a video using a Joint feature, which represents co-occurrence between multiple HOG features. Joint features represent the co-occurrence of the HOG features of two cells combined by the first-stage Real AdaBoost. Next, the generated Joint features are input to the second-stage Real AdaBoost, which constructs the final classifier. In this way, it is possible to capture shape symmetry and edge continuity, which single HOG features cannot do, so highly accurate detection is possible. We report experiments involving the detection of humans and vehicles performed to test the effectiveness of the proposed method. Furthermore, we also describe effectiveness of multi-resolution HOG representation, co-occurrence of spatio-temporal features, and co-occurrence of depth information obtained by a TOF (Time of Flight) camera.