Article Preview
Top1. Introduction
In recent years, the security of public places received much attention and there has been a growing trend to use video cameras for monitoring and surveillance purpose. Multiple human detection and tracking is a common problem in computer vision. Multiple human tracking systems finds its application in visual surveillance, intelligent living environment, event monitoring systems, anomaly detection / intruder detection, Motion Capture and Recognition, Traffic flow measurement, Accident detection on highways, Automotive safety, Intelligent control and human behavior analysis. Surveillance is performed in both outdoor and indoor environments. Surveillance for outdoor videos have some challenges like illumination variation, crowd, various types of objects, camera position, noise image due to a poor quality image source, varying pose, occlusions etc. are makes complex the system. Outdoor videos are obtained from cameras that are generally deployed outside buildings, by roads, in park and public areas, in fields etc. It is not necessary that humans will only be visible in the outdoor videos. Cars, motor cycles, buses, trucks, animals, birds etc. are also part of the other objects that are visible in the outdoor videos (Ansuman et al., 2014).
The person tracking was performed from offline data using background subtraction and multiple cameras. It introduced some delay in segmentation because sensor fusion is done by rendering foregrounds from multiple sensors image (John et al., 2000).
However a single camera can track the objects within the camera’s field of view. If the object goes out of the field of view, it is not possible to track the object. Continuous tracking of people over a wide area is much more powerful than in a single camera view because many interesting events happen over a large spatial/temporal domain. To track the objects continuously in wide areas, it is necessary to fuse the data of multiple video cameras (Vasuhi & Vaidehi, 2014).
Inaccurate detection of humans in such applications may increase the number of false alarms. Therefore, fuzzy logic has been used to model a robust background for object detection (Mahapatra et al., 2013). The required object recognition in the video is another important problem, for wide baseline matching and object recognition, both for specific objects as well as for category-level schemes is utilized (Tuytelaars & Mikolajczyk, 2008).
Tracking is performed in larger area and field of view (FOV) of multiple cameras. For the limitation of coverage of a single camera and the price of high definition cameras, large scale surveillance system, usually employs multiple standard video cameras. Each camera has a specialized field of view (FOV), and n FOVs make up the scene image of the whole long road just as resulted from one large view camera. Therefore, this paper proposes a system to track multiple objects continuously over wide area. The system uses multiple cameras with a little overlapping between the FOV’s. Common points are identified in the overlapped field of view and panorama view (Brown & Lowe, 2007) (He et al., 2010) has been created using stitching technique and to perform stitching, homography matrix is used.
The image features have many properties that make them suitable for matching differing images of an object or scene. The features are invariant to image scaling and rotation, and partially invariant to change in illumination and 3D camera viewpoint (Lowe, 2004).
The image alignment algorithms can discover the correspondence relationships among images with varying degrees of overlap. They are ideally suited for applications such as video stabilization, summarization, and the creation of panoramic mosaics. Image stitching algorithms take the alignment estimates produced by such registration scene movement as well as varying image exposures (Szeliski, 2006).
When multiple cameras monitor a scene, some regions may be occluded in one view but visible in other views. Different camera frames are warped and blended with respect to one reference camera. Tracking is performed in the enlarged stitched video and tracking across cameras can also continuously maintain the identity of a moving human (Rashimi et al., 2013) (Tao et al., 2005).
Fully integrated real-time system to track humans with a network of stereo sensors over a wide area has been presented by Tao et al. (2005). Detection and tracking is performed, frame to frame tracking of moving objects of stitched image obtained from multiple video cameras (Esra & Margrit, 2010). Kalman filter is employed to simultaneously track objects in 2D image coordinates and 3D coordinates for each camera. The 2D/3D trackers of each camera share information to improve the performance and trajectory prediction, which is used in case of occlusions (Black et al., 2002)[14].