Turning movement count data; that is, vehicle volumes broken down by movement, approach, and time period, are the foundation of signal performance evaluations and a crucial component of data-driven decision-making processes used by transportation agencies. Unfortunately, the availability of quality turning movement count data is arguably not the norm for agencies. In fact, the 2012 National Traffic Signal Report Card conducted by the National Transportation Operations Coalition identified traffic monitoring and data collection practices in the United States as weak by giving the practices an “F” grade. To this day, some of the current practices rely on manual procedures that limit the amount of data available. Automated methods can be temporarily installed at an intersection, but these are intended to improve on the traditional manual counts used and not to produce continuous count volumes. Permanent counting systems are unable to classify vehicles into their corresponding movements on shared lanes unless supplemental infrastructure is installed or additional count zones are defined. As part of this NCHRP IDEA Stage 1 project the research team has shown that an algorithm that produces turning movement counts reports using vehicle trajectory data extracted from existing vehicle detection infrastructure can be created. The algorithm developed in this NCHRP IDEA project differs from existing approaches in that it does not rely on count zones on exit approaches or the use of time stamps of detection calls. As a proof of concept, the algorithm has shown significant promise in terms of performance at typical intersections. While changes to the algorithm are needed, results obtained are encouraging and can be used by those familiar with data analysis and collection techniques. Innovation as part of this project can be classified in two different areas. First, showing that exploiting the capabilities of existing infrastructure is possible and, second, developing an analysis procedure for vehicle trajectory data. The capability of existing devices was shown by extracting vehicle trajectory data from an existing, commercially available, radar-based vehicle detection system that was already installed at a signalized intersection as a replacement for inductive loops. Vehicle trajectory data extraction was possible through a custom software program that taps into the underlying data stream of the radar-based vehicle detection system without interfering with the primary function of vehicle detection. The trajectory data were then analyzed using an algorithm implemented in the R programming language to classify the vehicle movements into left, thru, and right. The combined data collection and analysis approach is built on top of open source technology, is independent of the controller type, and could be deployed in numerous hardware platforms, thus eliminating the need for proprietary solutions or significant capital investments that often are part of projects focused on monitoring performance measures. Performance obtained can be considered better than that of some of existing technologies and can be measured in two different areas. First, at intersections with simple geometry such as the main site used for data collection (three lanes per approach, no channelization, and a homogeneous stop bar location across all lanes of each approach). At the aforementioned type of intersection, performance of the algorithm was measured at the count period and movement level. A count period was defined as a 15-minute interval for a specific vehicle movement. Vehicle volumes by count period produced by the algorithm were compared with volumes from a manual count obtained from video of the intersection. Turning movement count volumes produced by the algorithm had an average error of ?0.26 vehicles per 15-minute count period when compared with manual counts and an average absolute error of 2.31 vehicles. The performance of the algorithm was evaluated at what could be considered a non-typical intersection (five lanes per approach, one bike lane, and a non-uniform stop bar position on each lane). The mechanics of the algorithm performed as expected. However, the nature of the intersection in terms of geometry and traffic resulted in a lower than expected algorithm performance, especially for the left-turn movement located the farthest away from the radar-based vehicle detection system. In the case of the left turn, accuracy dropped to less than 50%, due to occlusion. A review of the scenarios under which significant accuracy reductions by count period were experienced reveals the impact that non-typical intersections have on the performance of radar detection systems and consequently on the algorithm. Therefore, the need to improve it further in order to handle special circumstances related to geometry and traffic characteristics is highlighted. As discussed, performance reported for the algorithm is measured not only every 15 minutes but also by individual movement. Therefore, the approach used by the team to report performance is one that goes beyond the standard practice of vehicle detection systems manufacturers in which performance is reported once a certain volume threshold is met and for entire approaches. This reporting approach was selected because it allows us to look at errors at a more detailed level and identify areas for future improvement. In order to successfully commercialize this innovation, the algorithm needs to be improved, a prototype data collection device developed that can be installed inside a signal cabinet, and a centralized software tool devised to manage multiple data collection devices. Although the results from the project are encouraging, algorithm improvements are needed in order to have a market-ready solution. Future work in terms of algorithm improvement should focus on making the algorithm capable of handling the following scenarios: * Different intersection configurations; the algorithm developed as part of the project provides satisfactory performance on what could be considered a textbook intersection. For example, no testing has been done on intersections with channelized lanes and testing is required at intersections with a non-uniform stop bar location. * Intersections with significant presence of heavy vehicles. Results from the second supplemental site show the need for further development in order to improve movement detection as a result of occlusion effects. * Intersections with two detection devices per approach. For intersections with 5+ lanes per approach more than one radar device is desirable and will require changes to the algorithm in order to properly merge and analyse multiple data sources. Modifications will require data collection across different locations in the country in order to obtain vehicle trajectory datasets from a wide array of geometric and traffic conditions. These modifications will in turn require further analysis and performance evaluations as part of an iterative development process. In terms of data collection, the system should be tested on a smaller hardware platform such as a development board and configured for an environment in which the data analysis algorithm and data collection system are on separate locations. (Author/publisher)
Abstract