Pedestrian Intention Detection for Autonomous Driving
Examensarbete för masterexamen
Software engineering and technology (MPSOF), MSc
Autonomous driving has drawn growing research interest in recent years, where considerable amount of software is being developed and integrated into vehicles. Pedestrian intention detection is one of the most important research topics in autonomous driving, since pedestrians are most common traffic participants, apart from vehicles themselves. This thesis addresses the problem of identifying the intention of pedestrians from a forward-looking camera mounted on a moving vehicle. In this thesis, we first studied existing work that is closely related to the addressed problem, i.e., pedestrian intention detection, and then proposed a complete workflow for solving the problem, which consists of three major steps: (i) pedestrian detection, based on a two-stage classification scheme using aggregate channel features (ACF) and histogram of oriented gradients (HOG); (ii) pedestrian tracking, based on Kalman filtering; and (iii) pedestrian intention detection. As a core part of the thesis work, pedestrian intention detection is formulated as a 3-class classification problem, where the intention of a pedestrian is classified to (a) staying off the road; (b) about to cross the road; or (c) crossing the road. This is achieved by extracting time-dependent features based on displacements of pedestrians, and taking into account contextual information from the detection of ego-lane. The proposed method was evaluated on 6 different videos containing representative scenarios that are selected from several publicly available datasets and one private dataset. Preliminary results have shown high detection accuracy (average 83.96%) and small false alarms (average 9.47%) for predicting pedestrian intentions, providing evidence on the effectiveness of the proposed method.
Data- och informationsvetenskap , Computer and Information Science