Stationary Detection of the Pedestrian?s Intention at Intersections

This paper focuses on stat ionary detect ion of the pedestrian's intention to enter the traffic lane at intersections. We use an Interacting Multiple Model Extended Kalman Filter (IMM-EKF) based tracking approach as basic method to recognize this. In addition, we propose a novel Motion Contour image based HOG-like descriptor (MCHOG) in combination with Support Vector Machine (SVM) classification that reaches the decision at an accuracy of 99 % within the initial step at the curb of smart infrastructure. MCHOG implicitly comprises the body language of gait initiation, especially the body bending and the spread of legs. As a result of a case study at laboratory conditions we present Receiver Operating Characteristic (ROC) performance data and an evaluation of the span of time necessary for recognition. While MCHOG in special cases indicates detection of the intention before the whole body moves, on average it allows for detection of gait initiation within 6 frames at a frame rate of 50 Hz and an accuracy of 80%. In addition to a comprehensive evaluation under laboratory conditions we demonstrate feasibility of the methods in a real world intersection scenario and we show the gain which we achieve with MCHOG compared to the IMM-EKF approach.