The unscented Kalman filter for pedestrian tracking from a moving host

Abstract

In this paper we present a time-efficient estimation framework for camera-based pedestrian tracking from a moving host car using a monocular camera. An image processing system processes the camera output to find the location of objects of interest in each frame. The position and sensor information about the host translation and rotation are passed to a tracking module. The module uses the position of the detected object's foot point as measurement input and connects them over time to estimate the movement of the objects of interest in order to reduce noise and single frame failures in the detection process. We have developed a new method to estimate the target movement which takes into account the host movement and allows to exploit prior information about the intrinsic and extrinsic camera parameters. The basic idea is to assume that host and target movements can be modelled as 2-dimensional movements on a flat ground-plane. Our developed motion model is based on this assumption and includes host motion as well as the target ego motion. A measurement is modelled as a perspective projection of a point on the groundplane to the image plane. The motion and the measurement model are combined by an Unscented Kalman filter. This filter is relatively new and has not been applied for pedestrian tracking before. Finally, we present a new logical initialization strategy for the selected filter, a part that is left out by most other publications. First results indicate that our approach gives good tracking results and allows to track pedestrians from a moving host in real time.

Will the Pedestrian Cross? A Study on Pedestrian Path Prediction

“…The same idea can be applied to KF extensions to either multiple linear dynamical models, e.g., the IMM KF [4], or to nonlinear models, e.g., the extended KF or the unscented KF (see [7] and [8] for applications to pedestrian tracking).…”

Context-Based Pedestrian Path Prediction

“…The state distribution of a LDS can be propagated into the future without incorporating new observations to account for missing measurements, or to perform path prediction. The Extended and Unscented KF [24] can, to a certain degree, account for non-linear dynamical or measurement models, but Switching LDS (SLDS) are needed for maneuvering targets that alternate various motion types. A SLDS uses a top-level discrete Markov chain to select per time step the system dynamics of the underlying LDS.…”

Pedestrian Path Prediction with Recursive Bayesian Filters: A Comparative Study

“…[15] use a UKF in a mono camera setup to track pedestrians on the ground plane (CV model). [19] apply the UKF to measurements from a stereo camera system comparing three different dynamical models (two CV and one constant position (CP) model) where two models have a state space in world coordinates and one in image coordinates.…”

“…[3,5,9,15,18,21] do not include any such evaluation. [2] provides accuracy figures only related to its KF approach in indoor setting.…”