Measurement of Pedestrian Traffic Using Feature-based Regression in the Spatiotemporal Domain

Measurement of pedestrian traffic in public areas (e.g., stations, airports, shopping malls or complex buildings) provides valuable information. For safety management, congestions can be detected to prevent accidents in their early stage by monitoring the pedestrian traffic continuously. Knowing the number of people working in a large building may help to when designing evacuation plans. For marketing purposes, value assessments of shopping areas can be achieved based on traffic data because higher pedestrian traffic is directly linked to more sales. In building management, pedestrian traffic data can be utilized to optimize the number and working hours of staff. Power savings can be achieved by adjusting airconditioning and heating based on pedestrian traffic. Over the last decade, various computer vision methods have been studied to automatically measure the pedestrian traffic. One popular approach to pedestrian traffic measurement is the use of top-view cameras. In this approach, a camera is mounted vertically at the top of a gate or over a region of interest. Because of the superior viewpoint of the camera, pedestrians do not obscure each other in video frames. Hence the problem of pedestrian traffic measurement may be solved easily by detecting moving objects using foreground segmentation and tracking the detected blobs (Sexton et al., 1995; Kim et al., 2003). However, these methods fail when a number of people move close or slightly touch each other creating a single blob. Chen et al. resolved this problem by comparing the area of detected moving object with the area of one person to estimate the number of people in the blob (2006). Velipasalar et al. employed two-level hierarchical tracking to deal with pedestrians of complex movements interacting with each other (2006). Pedestrian traffic also can be studied by detecting humans using standard surveillance cameras that do not require a specific viewpoint. Similar to top-view camera based methods, some of these methods perform foreground segmentation to distinguish moving objects. However, for oblique camera angles, multiple pedestrians easily appear as merged blobs in a video frame. The detected foreground blobs are segmented into individuals by modeling humans as ellipsoids (Zhao et al, 2004, 2008) or rectangles (Liu et al, 2005; Beleznai et al., 2006) cooperating with the known camera geometry. Based on based on their shapes and apperances, humans can also be detected directly from image frames without separating out foreground blobs. Viola et al. detected humans using appearance and motion information