自動車の衝突事故の回避には、自車両周辺の道路ユーザーを検出するのみでなく、その行動を理解し、予測することが必要です。 当研究では、歩行者と自転車に焦点を当て、車載カメラからの行動解析のフレームワークを提案します。 提案手法は、まず複数の特徴量を用いて歩行者と自転車を識別し、その後その頭部/胴部姿勢を推定します。 ここでは、従来の教師学習の代わりに半教師学習を用い、また頭部と胴部の姿勢の間の物理的制約と時系列的制約を考慮することで、 画像シーケンスにおいて正当かつ安定した推定を実現しています。

Collision avoidance systems are not only required to detect road users around vehicles, but also expected to understand and predict the behavior of road users for risk assessment. We focuses on two kinds of similar road users, pedestrian and cyclist, and proposes a behavior analysis framework from on-board camera. The proposed method firstly recognizes the type of road user, and then estimates the pose of road user. The first recognition phase employs a cascade structured classifier. This classifier distinguishes cyclist from pedestrian using multiple features and discriminative local area, in order to achieve a high recognition rate. In the second pose estimation phase, both head orientation and body orientation are estimated. In order to obtain more accurate classifier, Semi-Supervised Learning is applied instead of the conventional Supervised Learning method for training. Moreover, the human physical model constraint and temporal constraint are considered, which assist the pose estimation to produce reasonable and stable result in video sequence. A series of experiments demonstrate the effectiveness of the proposed method.