Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
79 C h a p t e r 1 3 The aim of this project was to develop and validate a new obser- vation tool for vehicle safety researchâan automated video tracking system, capable of wide-scale use, to captures vehicle trajectories, find traffic conflicts, and compute metrics such as time to collision. The concept design and system devel- opment took place with reference to the research agenda of SHRP 2. The system has been shown capable of high-fidelity tracking of vehicles for exposure and conflict analysis, and addressing research questions that are not amenable to other techniques. It achieves this by detection of vehicles using a small number of clusters, localizing the vehicle using overlap- ping polygons, and then using dense cluster sets to track each vehicle (see Figure 13.1). The system is immediately capable of supporting new research into risk factors and countermeasure performance. This is especially important in the areas where vehicle-based naturalistic data collection is silent, such as turning conflicts at intersections. This report has considered in detail the design and develop- ment of the Site Observer. No comparable system currently exists, and the outcome of this project has been to develop a robust prototype system, together with supporting software and sample data from a small field trial; the system has been designed, built, and implemented and has captured represen- tative data. The system was designed to address key safety research ques- tions posed by SHRP 2. In particular, it is capable of deter- mining a variety of conflict metrics and provides a prototype working tool for evaluating safety performance of highways in the future. The project focused on turning conflicts at inter- sections, but the system is fully capable of being used at other locations and for other conflicts of particular interest. The sys- tem and the data it provides go beyond anything currently available, and the system is expected to become an invaluable tool for researchers and practitioners. The system uses multiple machine vision cameras, net- worked together and accurately synchronized, and is orga- nized in a hierarchical way with early stage processing taking place on site, local to where each camera is installed. In the future, it will be possible to integrate the image processing algorithms into the camera; such cameras with built-in digi- tal signal processors are now available on the market. Cou- pled with wireless data transfer, the design has the scope to be easily installed and highly portable, reducing installation and setup times. Another way that setup time can be reduced is to extend the process of on-site camera calibration to the point where the lidar survey is not required. This will also reduce cost on a per- installation basis, although for flexibility a pan-tilt-zoom unit should be considered. The whole calibration process may then be performed on site using a reference vehicle fitted with a highly visible marker or set of markers. Software exists to per- form the calibrations automatically, provided the vehicle (and thus the markers) are located with sufficient accuracy and a sufficient number of traces are found. Real-time kinematics analysis can be performed after collecting raw GPS data to obtain centimeter-level positioning accuracy, and provided the vehicle is driven slowly to avoid excessive body roll or suspension travel, the necessary accuracy can be obtained. Additional survey points need to be located to determine an intersection map; fixed markers will also be needed to main- tain the calibrated camera position or determine any small off- sets during use. The requirement will again be met that vertical curvature of the roads is accounted for during vehicle track- ing; the modified approach will be equivalent to the use of lidar survey data, avoiding the large data volumes and reducing setup time. A major advantage of the Site Observer is its ability to reduce data volumes and offer full automation. The system does not require special high-level camera locations, although cameras do need to be mounted strategically to maximize overlapping fields of view. It has been found that covering all vehicle posi- tions with several camera fields of view is important for robust and accurate positioning. Synthesis of extracted features has been carried out using database methods, followed by the Conclusions
80 and that these were simply not observed. However, it seems reasonable to conclude that again the proportion of missing vehicles typically is very low. It may be possible in future work to make this part of a formal analysis by using a physical trig- ger, such as a loop detector, for comparison. Even if both the video and reference system miss a small proportion of actual vehicles, provided the sources of missed trajectories are inde- pendent, it will be possible to put a firm estimate of the pro- portion of missed trajectories. Several aspects of the system allow scope for optimization and further refinement: the low-level image processing code can be made more efficient for real-time use, camera calibra- tion can be automated using identifiable moving point targets on site (e.g., on the roof of a calibration vehicle), and greater use of wireless networks can allow researchers to control and monitor the system remotely. Also, the current version of the Site Observer has not fully exploited all information available in the feature sets. For example, the use of overlapping blobs projected onto the ground plane has been applied only at selected time instances, and their use can be increased. Overall, there is a tradeoff between computational efficiency and accu- racy of results, and through additional field trials, at sites with a greater density of traffic, available tradeoffs can be exploited. The current version of the Site Observer only sees cars and trucks, and this is based on the design intent within the project; for example, it is blind to bicycles and pedestrians. But the methods are more widely applicable; tracking and localiza- tion are feasible provided the initializing detection (triggering) mechanisms are set up for detecting these other road users. The same basic feature sets can be used, although camera posi- tioning should take these extended requirements into account. Motion variables were compared with those obtained from test vehicles. It was shown that velocities and accelerations can be estimated and provide useful information. However, it is not feasible to use the Site Observer to replicate what is observed in a naturalistic vehicle-based field study, quite apart from the fact that drivers are anonymous to the system. Site-based data are complementary to vehicle-based data, and there are opportunities to join these kinds of data in the future; when instrumented vehicles drive through an observed site, the data can be joined retrospectively by relating time and position information. This approach may offer new opportu- nities for addressing complex safety questions when driver factors are considered crucial. The report has shown what the system currently does and mentions future visions for a more portable and packaged sys- tem, in effect a consumer product for professional researchers and practitioners. In the meantime, the system as it exists can be usefully deployed in a number ways. With current soft- ware and hardware, it can be installed at new locations where traffic densities and conflict levels are more challenging; this could be at a signalized intersection or really at any location application of 3-D mappings and dynamic analysis to create time histories of position, velocities, and accelerations. Implementation took place at a suburban intersection in Ann Arbor, Michigan; approximately 17,000 vehicle trajec- tories were captured. A feature database was constructed, trajectories were extracted, and sample conflict analyses were undertaken for path crossing and rear-end conflict types (LTAP/OD and RTIP). The current system uses four cam- eras and five computers but is limited only by hardware costs; the design is fully scalable to larger installations. An important contribution, beyond the actual development and use of the system, has been to link requirements and design principles directly to the needs of the SHRP 2 Safety program. The project has also validated the new design concept based on a modular and layered approach. While there remains scope for optimization and refinement, the systemâs tracking accu- racy exceeds that of comparable research systems, while at the same time decoupling the image analysis from the detection and tracking analysis. To achieve this duality, two comple- mentary feature typesâclusters and blobsâhave been required Neither feature type on its own is sufficient for the purpose. The main purpose of the system is to capture research data. It is not essential that all trajectories are captured, only that loss of data does not bias results. Every time a system is switched on, the preceding data has been lost, and when it is switched off, more data are lost. The team samples only the population of interest. Data loss is not the vital issue; the goal is to collect sufficient unbiased data to answer safety research questions. It has been mentioned that the proportion of cor- rupt vehicle trajectories at the test site was found to be around 1%, which is certainly too small to adversely affect the valid- ity of safety research results. The proportion of missing vehicle trajectories is a more difficult issue, and again no firm num- bers were determined. In manual video review, from many tens of hours of viewing vehicles, together with overlays of triggered cluster tracks, no cases were found in which a vehicle failed to show at least one such track. This was not part of a formal analysis or controlled experiment, so it is possible that certain (unknown) conditions can lead to missed vehicles Figure 13.1. Motion capture by multiple cluster tracking.
81 general, relating conflicts to crashes at a single site may have limited scope, unless crashes are sufficiently frequent there. To overcome this limitation, one research option involves recording conflict data at a single location or at a small num- ber of locations. Then, relating conflicts to turning counts (and other factors such as time of day), it will be possible to impute conflict rates at many other comparable locations for which turning count data are available. The resulting con- flict and exposure rates can then be directly related to a larger pool of crash numbers and types. Such an analysis will be of immediate benefit in improving the understanding of risk factors and the relationship between crashes and conflicts. of particular interest. The purpose will be to relate conflicts to crashes, either directly or with reference to other variables such as detailed turning counts. Useful data can be collected in a matter of weeks; it is not necessary to wait for crashes to occur. This makes for a very powerful research and evalua- tion tool. Future research At the field test site used, the number of historical crashes was seen to be very low, but when the many thousands of such sites are considered, the safety problem is far from trivial. In
