Object-based Detection and Classification of Vehicles from High-resolution Aerial Photography

Vehicle counts and truck percentages are important input variables in both noise pollution and air quality models, but the acquisition of these variables through fixed-point methods can be expensive, labor-intensive, and provide incomplete spatial sampling. The increasing availability and decreasing cost of high spatial resolution imagery provides an opportunity to improve the descriptive ability of traffic volume analysis. This study describes an object-based classification technique to extract vehicle volumes and vehicle type distributions from aerial photos sampled throughout a large metropolitan area. We developed rules for optimizing segmentation parameters, and used feature space optimization to choose classification attributes and develop fuzzy-set memberships for classification. Vehicles were extracted from street areas with 91.8 percent accuracy. Furthermore, separation of vehicles into classes based on car, medium-sized truck, and buses/heavy truck definitions was achieved with 87.5 percent accuracy. We discuss implications of these results for traffic volume analysis and parameterization of existing noise and air pollution models, and suggest future work for traffic assessment using high-resolution remotely-sensed imagery. Introduction Accurate vehicle counts and truck percentages are important for traffic volume analysis, and serve as input variables to both noise pollution and air quality models. For example, vehicle traffic counts are directly linked to noise-related health impacts in urban environments (Seto et al., 2007). In order to calculate noise volumes for a given highway or street segment, noise pollution models such as the Federal Highway Administration Traffic Noise Model (Menge et al., 1998) require traffic volume and vehicle-type percentages along each street segment. Likewise, traffic volumes, vehicle type distributions and vehicle densities are important inputs to models of urban PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING J u l y 2009 871 Ashley C. Holt and Peng Gong are with the Department of Environmental Science, Policy, and Management, College of Natural Resources, University of California, Berkeley, CA ([email protected]). Edmund Y.W. Seto is with Environmental Health Sciences, School of Public Health, University of California, Berkeley, CA. Tom Rivard is with the Department of Public Health, San Francisco, CA. Photogrammetric Engineering & Remote Sensing Vol. 75, No. 7, July 2009, pp. 871–880. 0099-1112/09/7507–0871/$3.00/0 © 2009 American Society for Photogrammetry and Remote Sensing Object-based Detection and Classification of Vehicles fr om High-resolution Aerial Photography Ashley C. Holt, Edmund Y.W. Seto, Tom Rivard, and Peng Gong vehicular emissions (Lyons et al., 2003); these models, in conjunction with pollutant dispersion models, aid in the assessment of human exposure and overall air quality (Kunzli et al., 2000). In addition to serving as important input variables to both noise pollution and air quality models, accurate vehicle counts also have application to transportation planning, traffic flow studies, pavement maintenance programs, environmental justice studies (Forkenbrock and Schweitzer, 1999; Forkenbrock and Weisbrod, 2001) and the promotion of pedestrian and bicycle safety (Wachtel and Lewiston, 1996). There are several different methods, involving both direct and remote sensing, which can be used to detect vehicle traffic volumes along a given street segment. Technologies such as pneumatic tube systems are portable and can provide constant monitoring along a street or highway segment; on the other hand, this method is labor-intensive, involving the installation and subsequent removal of equipment for each monitoring location and time period, and in most cases does not provide the capability of detecting different vehicle types or sizes (Bellemans et al., 2000). An alternate option is the use of remotely sensed counts derived from fixed-point counters, such as loop counters or video monitoring. For example, video and infrared monitoring can provide constant-stream information about traffic flows along selected streets (Graettinger et al., 2005; Pless and Jurgens, 2004). One drawback with fixed-point counters, however, whether they involve direct or remote sensing technologies, is that sampling may be limited and is often conducted reactively, that is to say, on streets with known high traffic volumes or with pre-existing traffic or pedestrian-safety issues. Over large metropolitan areas with complex traffic patterns, detection technologies which provide more complete sampling capabilities can add an additional layer of traffic information to data acquired from direct or fixedpoint sensors. The increasing availability of affordable aerial photography and high-spatial resolution satellite imagery will provide new opportunities for deriving Annual Average Daily Traffic (AADT) (McCord et al., 2002). In addition, these data could contribute to spatiallycomprehensive assessments of noise and air pollution model parameters, such as traffic counts, as well as car, truck, and urban bus percentages. 871-880_07-079.qxd 16/6/09 2:36 PM Page 871