[download id=”4″ format=”4″]
We present a highly effective solution for multiple target tracking in sequences of images taken from a single rotating and zooming camera.
Tracking of moving targets is performed in world coordinates using an estimate of camera pose and focal length at each frame. Estimation of target position and exploitation of geometric relationships between the current view and the 3D imaged world allow improvement in accuracy of target initialization, in the quality of the association of measurements to targets and therefore in the overall tracking performance.
Probabilistic data association based on target motion is employed to avoid degradation of tracking. We present an extensive set of experiments with a novel dataset containing sequences taken from a rotating and zooming camera under different conditions. Validation of the method, in comparison with other state of the art approaches, is also shown on standard sequences taken from both PTZ and fixed cameras.
Sequences from a PTZ camera
- Seq.#1 “Long”: taken outdoor from a Sony SNC-RZ30 camera. The sequence has low target density with tracks of length up to 70 meters. The PTZ camera is steered to follow targets while continuously zooming-in;
- Seq.#2 “Focus”: taken outdoor from a Sony SNC- RZ30 camera. This sequence has medium target density and covers approximately the same distance as Seq.#1. The camera zooms on a target for a few frames, then it returns to the group of targets. Sunlight results in strong variations in target appearance;
- Seq.#3 “Dense”: taken outdoor from a Sony SNC- RZ30 camera. This sequence has the highest density of targets. Targets have small size and move erratically, entering and leaving the field of view, and experience frequent occlusions. They have similar appearance in most cases;
- Seq.#4 “Rapid Motion”: taken indoor from an Axis Q6032-E camera. Many targets move simultaneously in this scene while the camera performs fast patrolling and zoom-in operations. Targets partially occlude each other frequently and also exit from the field of view for a period before re-entering.
Video Results from a PTZ camera