A camera vision-based collision avoidance system (CAS) installed on buses in Miami-Dade increased transit operator response to rear-end and pedestrian conflict scenarios, but had mixed results with respect to user acceptance.
Florida DOT commissioned an assessment of the effectiveness of collision avoidance systems installed on-board transit buses operated by the Miami-Dade Public Works Department.
Made Public Date


Miami-Dade County
United States

Testing of a Vision-Based Pedestrian Collision Warning System on Transit Vehicles

Summary Information

Florida DOT, in collaboration with Miami-Dade County Public Works Department, conducted an assessment of a vision-based collision warning system for transit buses. Several key scenarios of interest, including transit vehicle-pedestrian and rear-end conflicts, were assessed by comparing driver responses when the system was active, and inactive but recording. In the rear-end scenario, operators were observed to drive more conservatively when the system was active, resulting in a reduced frequency of hard braking events (once every 166 miles vs. every 105 miles, on average). With the system active, overall yielding behavior to pedestrians increased from 58 percent to 88 percent.


The project installed a camera-vision collision warning system on board ten transit buses using a randomized experimental design. Five buses acted as a control group, where the system was silent (“stealth mode”) and did not present alerts or warning to the bus operator. In the other five buses, the system was active and presented alerts to the driver as designed. Drivers were assigned randomly to each group. Several evaluation objectives were planned with corresponding performance measures:

  • Objective 1: Assess the system effectiveness in reducing rear-end crashes
  • Objective 2: Assess the system effectiveness in reducing pedestrian and bicycle crashes
  • Objective 3: Assess the user acceptance of the system
  • Objective 4: Assess the ease of the installations and operations of the devices
  • Objective 5: Assess the cost-effectiveness of the technology.


The project analyzed three years of historical incident data for the Miami-Dade Transportation and Public Works buses to determine potential benefits from avoided crashes. The analysis used records from 5,164 collisions and determined that 38 percent of crashes occurred at intersections, and 28 percent occurred near bus stops. At least 15 percent of crashes resulted in injuries, and 38 percent were associated with inappropriate bus operator actions. Route specific data was also used to assess the distribution of potential crashes to be prevented by collision avoidance systems.

The project team analyzed the system performance based on recorded data and a sampled video confirmation for alerts. For example, recorded data streams were used to compare the time of alerts and the time the bus speed started decreasing, if there was an operator reaction. Video data were used to observe the external environment such as presence of actual conflict or not, and the yielding behavior to pedestrians.

To measure driver acceptance, a survey of the operators was conducted to capture perceptions of ease of use, usefulness, warning effectiveness in each scenario, and system accuracy.

Data Limitations

During the study, there were route changes that affected the ability to assess early data using the original experimental design. As a result, most of the analyses used the period after the route changes. With respect to driver feedback, a focus group was planned to gather in-depth information, but these data were not available at the time of the report. User acceptance results were based only on the survey responses.


Analysis of the recorded driver responses under the treatment and control conditions were compared. In some scenarios, such as headway warning and forward collision warning, only minor differences in response were observed (2-3 percent higher). However, for the urban FCW scenario, there was a significant increase in response with the system active, from 82 percent to 100 percent. In addition, the analysis of frequency of hard braking events found that with the system active, hard braking occurred on average every 166 miles, as compared with every 105 miles when the system was inactive. In the scenarios where the system alerts the driver of pedestrians, the analysis found that the frequency of yielding was 88 percent when the system was active versus 58 percent when the system was in stealth mode.

Alerts were also assessed based on a sample of video analysis to determine whether the issuance was consistent with the scenario-specific thresholds. In several scenarios, reductions in the false alert rate were observed due to a change in route, based only on the analysis of stealth mode data. For example, in the forward scenarios, false alerts were 20 percent , 38 percent, and 18 percent in the first route and 2 percent, 2 percent, and 18 percent after the route change for headway warning, urban forward collision warning, and forward collision warning, respectively. In pedestrian scenarios, the right-rear and left-front false alerts were 45 percent and 40 percent before route change and 10 percent and 13 percent after. There were not clear reasons reported for these effects associated with the route change.

Operators had varied opinions of the system usefulness. Overall, 22 percent of drivers thought the system was useful while 57 percent disagreed. Similarly, 55 percent of drivers would not recommend the system to other drivers while 18 percent would recommend it. These proportions were similar to the perceived alert effectiveness and accuracy. Overall, 20 percent of respondents thought that the system was accurate while 54 percent disagreed.