Multi-camera tracking for airport surveillance applications
D. Thirde, Mark G. Borg, James M. Ferryman, Josep Aguilera, Martin Kampel, Gustavo J. Fernández, J. Thirde · CentAUR (University of Reading) · 2006
Abstract This paper presents the multi-camera trackingcomponent of a complete surveillance system that was de-veloped as part of the AVITRACK project. The aim of theproject is to automatically recognise activities around aparked aircraft in an airport apron area to improve the ef-ciency, safety and security of the servicing operation. Themulti-camera tracking module takes as input per-cameratracking and recognition results and fuses these into ob-ject estimates using a common spatio-temporal co-ordinateframe. The multi-camera localisation and tracking of ob-jects is evaluated for a range of test data. 1 Introduction The aim of the EU AVITRACK project is to automaticallyrecognise activities around a parked aircraft in an airportapron area to improve the efcienc y, safety and security ofthe operation. A combination of visual surveillance andevent recognition algorithms are applied in a decentralisedmulti-camera end-to-end system providing real-time recog-nition of the activities and interactions of numerous vehi-cles and personnel in a dynamic environment. A main re-quirement behind the adopted architecture is that the imple-mented system must be capable of monitoring and recognis-ing the apron activities over extended periods of time, op-erating in real-time (colour PAL at 12.5 FPS). In this paperthe multiple-camera object tracking system is discussed, theoutput of this system is a set of estimated objects on theapron.The tracking system discussed in this paper comprisesper-camera (2D) video frame tracking and multi-camera(3D) fused object tracking. Video frame tracking methodsgenerally require methods to detect and track the objects ofinterest. Motion detection methods attempt to locate con-nected regions of pixels that represent the moving objectswithin the scene; there are many ways to achieve this in-cluding frame to frame differencing, background subtrac-tion and motion analysis (e.g. optical o w) techniques. Im-age plane based object tracking methods take as input theresult from the motion detection stage and commonly ap-ply trajectory or appearance analysis to predict, associateand update previously observed objects in the current timestep. The tracking algorithms have to deal with motion de-tection errors and complex object interactions in the con-gested apron area e.g. merging, occlusion, fragmentation,non-rigid motion, etc. Apron analysis presents further chal-lenges due to the size of the tracked objects with prolongedocclusionsoccuringfrequentlythroughoutapronoperations.To recognise the tracked objects in the observed scene bothtop-down(e.g. [8]) and bottom-up methods(e.g. [6]) can beapplied. The challenges faced in apron monitoring are thequantity (28 categories) and similarity of objects to be clas-sied e.g. many vehicles have similar appearance and size.Multi-camera tracking combines the data measured by theindividual cameras to maximise the useful information con-tent of the observed apron. The main challenge for apronmonitoring is the tracking of large objects with signicantsize, existing methods generally assume point sources [3]and therefore extra descriptors are required to improve theassociation. People entering and exiting vehicles also posea problem in that the objects are only partially visible there-fore they cannot be localised using the ground plane.The AVITRACK Scene Tracking module is responsiblefor the estimation of objects on the airport apron in twodistinct stages S per camera (2D) object tracking and cen-tralised world (3D) object tracking. The per camera trackingsub-module (described in Section 2) consists of motion de-tection to nd the moving objects, followed by tracking inthe image plane of the camera. The tracked objects are sub-sequently classied using a hierarchical object recognitionscheme. The multi-camera tracking sub-module (detailed inSection 3) fuses the tracking results from the eight camerasinto object estimates. Finally, Section 4 gives experimentalresults showing the localisation and tracking performance.