Refining Correctness of Vehicle Detection and Tracking in Aerial Image Sequences by Means of Velocity and Trajectory Evaluation

Derivation of statistical traffic data is highly dependent on the balance of detection and false alarm rates. In case false alarms have not been eliminated in the initial detection phase, they are often subsequently tracked, though, resulting in trajectories that do not match the true traffic situation. This finally leads to derivation of erroneous traffic parameters within the individual road segments. In this paper, a method is described how to eliminate false alarms by evaluating the trajectories and velocities of a tracking procedure. Basically, two types of false alarms are considered which bias the statistics of traffic data: The first type deals with redundant detections that lead to multiple trajectories biasing the statistics. The second type comprises false alarms that belong to the static background inducing zero-velocity into the statistics. We show that the presented procedure is able to increase the total correctness of detection and tracking from 65% up to 95% which allows a much more precise calculation of traffic flow parameters. 1. TRAFFIC MONITORING The task of collecting wide area traffic parameters plays important role in today’s traffic management. Aerial images offer a complement source to common measurement systems like induction loops and stationary video cameras. Besides giving a visual overview, image sequences which cover large areas can deliver a time snapshot of a spatially fully covered traffic situation of the recorded region. In recent years, traffic monitoring using airand space images became more and more attractive mainly due to the availability of cost-effective and flexible high-resolution systems mounted on aircrafts, i.e. the LUMOS/ANTAR system for traffic monitoring (Ernst et al., 2003; Ernst et al. 2005; Ruhé et al., 2007) or the 3K camera system (Kurz et al., 2007), or on HALE platforms and UAVs as presented in the Pegasus project (Everaerts et al., 2004). An extensive overview on recent developments is given, for instance, in (Stilla et al., 2005; Hinz et al., 2006; Lenhart et al., 2008). The following methods are especially designed for traffic monitoring with DLR’s 3K camera system. This system is able to capture image sequences with a frame rate of approx. 3Hz – 7Hz depending on the imaging mode (continuous imaging or bursts) with a spatial resolution of 20cm – 50cm depending on the flight height. Concepts for deriving traffic data from these aerial image sequences have been proposed in (Rosenbaum et al. 2008) and (Lenhart et al. 2008). The traffic parameters which are calculated from image sequences are namely the mean velocity and traffic density per road segment. The resulting parameters are then integrated into traffic flow models such as the DELPHI traffic portal illustrated in (Behrisch et al.). 2. INFLUENCE OF FALSE ALARMS Detection methods as proposed in (Rosenbaum et al. 2008) or (Lenhart et al. 2008) deliver a detection quality of about 60% completeness and 65-75% correctness. False alarms are mainly caused by structures which appear similar to vehicles, like i.e. belonging to shadows, road banks etc. The influence of the false alarm rate on the calculation of generic traffic parameters can be studied using, e.g., MonteCarlo simulations. In the following experiment a dense traffic scenario on a multi-lane highway was captured with an image sequence and all car trajectories were manually measured in this sequence, eventually leading to mean velocity profiles for each lane of the highway. Then, a predefined percentage of detections were selected at random positions along the road and contaminated with a specific percentage of random false alarms. Based on these data the velocity profiles were calculated for each lane again and compared to the reference data. As the estimation of the velocity profile depends strongly on the randomly selected positions of the cars, these experiments have been carried out 10000 times, in order to gain a certain statistic about the quality of the estimated profiles. The following table summarizes the RMS values and standard deviations for the estimated velocity profiles depending on the respective detection and false alarm rate. 50% detection rate 5% false alarm rate 50% detection rate 10% false alarm rate 50% detection rate 25% false alarm rate