Florida DOT has been using inductive loops and sensors at I-95 ramp signals in Miami-Dade County, but existing detectors only report volume and occupancy data. In order to support performance monitoring of ramp signals, researchers investigated the feasibility of using a video detection system for collecting performance monitoring information, such as queue length and waiting time. A three-step study was conducted to identify potential video detection systems for evaluation, conduct field testing of three systems to assess performance, and select a system for deployment based on several factors. In addition, the project team developed a web-based system for dynamically visualizing performance data from detector and incident data.
Performance measures used in the field test of three detection systems included detection accuracy for vehicle occupancy, and queue lengths based on different scenarios with different gap sizes and percentages of activated zones (i.e., zones that detected a vehicle present). The data collection period for initial field testing covered both late afternoon and evening, to include varied lighting conditions. The data gathered consisted of recorded video and measures including:
- Stop line traffic volumes at 5 and 15-minute intervals
- Stop line occupancy (percent) at 20-second intervals
- Maximum queue length (feet or number of vehicles) within 5-minute intervals
- Average queue length (feet or number of vehicles) at 5-minute intervals
- Average travel time on ramp or alternative measure of wait time, if available
The video recording was used as ground-truth data to assess the accuracy of the test data, and was supplemented by backup camera video and placement of traffic cones to facilitate manual data collection. Stop line occupancy was also compared with loop detector data from Florida DOT. Mean Absolute Percentage Error (MAPE) was the measurement chosen to represent an overall average, based on the percentage difference between the video system and ground truth data within each time interval.
- Based on field testing the three systems, the researchers discovered that the results varied in performance (accuracy) under different lighting conditions. For the first system, performance in counting was more accurate in daytime conditions as compared to nighttime, where vehicles tended to be undercounted. Overall, the MAPE was calculated at 11.6 percent. The second system also experienced good count performance in daytime conditions but experienced overcounting of vehicles during nighttime conditions. Overall, the MAPE for the second system was calculated to be 19.7 percent. The third system utilized a camera with thermal-sensing capabilities and had good performance in both day and night lighting conditions, resulting in a MAPE of 3.7 percent overall.
- Field testing also enabled researchers to determine that the capabilities (available performance data outputs) demonstrated by each system also varied substantially. Both of the first two systems only provided vehicle count data and video recordings, despite additional performance data outputs being specified in the field test plan.
The research team further assessed the other measures only available with the third (thermal) video detection system, coupled with the existing ramp signal software at the Traffic Management Centers (TMC):
- Vehicle occupancy data from the video detection system closely matched the data recorded in the loop detector database. After adjusting for differences in detector size, the MAPE was calculated to be 5.9 percent.
- The system was able to measure average vehicle queue lengths that matched well with those estimated from manual measurement, within a MAPE ranging between 24.2 and 31.4 percent. Based on the different scenarios studied, researchers considered this a reasonable range given the uncertainties associated with measuring moving ramp queues.
- The system similarly provided good estimates of the maximum vehicle queue lengths. The MAPE ranged between 6.2 and 8.4 percent across the different scenarios studied.
- The system performed equally well under both daytime and nighttime conditions.
- The system demonstrated that it could combine data from multiple cameras for vehicle queues that are longer than what one camera can cover.