2.1. Instruments
The methodology scheme is shown in
Figure 1. The observational investigation was performed using two cameras for field data collection. On the one hand, data related to the static geometric parameter data of the study segments and the dynamic traffic performance of non-motorized vehicles were obtained. On the other hand, naturalistic videos from a static camera were recorded to obtain the characteristics of each rider.
Diversified methods can be used for traffic flow data surveys. Krajewski et al. compared the drone-based approach with existing measurement methods (drone, infrastructure sensors, vehicle with series-production sensors, vehicle with highly automated driving sensors) in terms of the five requirements (naturalistic behavior, static scenario description, dynamic scenario description, effort effectiveness, and flexibility) [
29]. The comparison showed that drones have several strengths in terms of naturalistic behavior and static and dynamic scenario description. Unmanned aerial vehicles and cameras have recently been used in traffic safety studies including driving behaviors and riding behaviors [
30].
Given the research objective of this paper, it is necessary to obtain the overtaking phenomena of non-motorized vehicles on bicycle lanes. In this survey, the DJI Drone Inspire 2 is used with an external 360° pan/tilt camera and built-in SSD storage space, which supports 4K ultra-clear video recording (30 frames per second). With its flexible maneuverability, strong hover shooting stability, support for precise hovering without GPS, and satisfactory image anti-shake effect, the aircraft can still shoot stable videos even when flying in large movements. See
Figure 2 for the drone equipment and aerial photography site. In this study, it is also necessary to identify the individual attributes of riders, which cannot be clearly identified by shooting at high altitudes. Therefore, a video camera with a tripod set up on the roadside is used to capture the individual information of riders, as shown in
Figure 3.
2.2. Data Acquisition Procedures
In this paper, the overtaking behavior of non-motorized vehicles in Hefei is taken as the research object. Hefei is the capital of Anhui Province and is a typical large city located in eastern China. The city’s orientation, traffic characteristics, geometric road design, and riding behavior are similar to those in other major Chinese cities. By 2021, the number of electric bicycles in Hefei alone had reached 3 million [
31]. Like other cities in China, Hefei has high levels of non-motorized traffic and frequent overtaking during peak hours.
Before the official start of the traffic investigation, the shooting time and place of the drone should be determined. To reduce the adverse effect of strong sunlight on the later-stage speed extraction, video capture is carried out in the evening peak hours (17:30–18:30) with fine weather. In the selection of the survey site, the following principles are followed: the bicycle has a large flow and is a relatively continuous form of exercise; there is no influence of a large canopy, telephone pole, street lamp, or other obstructions above the non-motorized vehicle lane; the flight area meets the basic requirements of drone take-off, avoiding restricted areas such as military areas; and the drones and investigators cannot adversely affect the bicycle traffic flow. To reduce the interference of motor vehicles and pedestrians on the bicycle flow, the survey segments selected in this study are all in the form of machine non-isolation. In the early stage of the pre-investigation, it was found that the road segments with isolated markings tend to have a small bicycle flow. Even if bicycles pass through, riders do not drive in the markings, and they have a high probability of occupying motor vehicle lanes. Therefore, the road segment that is physically isolated between motor vehicles and non-motorized vehicles is finally selected as the drone video acquisition point. See
Table 1 for specific information.
There were three members of the research team at the scene: one member controlled the drone, one took care of the roadside camera, and one took on the auxiliary work. The researchers first debugged and prepared the equipment, set up the roadside camera, and adjusted the angle when the drone reached a reasonable height. Then, the video recording buttons of the drone and the roadside camera were pressed at the same time to ensure that the video taken by the drone and the video taken by the camera corresponded to each other during the later data-processing stage. After the shooting, the researchers measured and recorded the width, length, and other data from the survey road segment.
2.3. Data Extraction and Processing
The speed data are extracted using Simi Motion software and manual punctuation. The steps are as follows:
Step one: Video import. The video taken by the drone is converted into a format readable by Simi Motion software 9.2.1 (Simi Reality Motion Systems GmbH, Munich, Germany) by format factory software, and then imported into the Simi Motion software.
Step two: Image calibration. According to the measured distance in the traffic investigation, the pixel distance of the video is calibrated. Using the four-point calibration method, four points of 1, 2, 3, and 4 are selected as marking points, and the actual measured distances of 1–2 and 2–3 are input to complete the image calibration of the video, as shown in
Figure 4.
Step three: Non-motorized vehicle tracking. Manual dotting is used to mark the center point of a non-motorized vehicle as a tracking mark point. If the mark point is separated from the vehicle in the process of automatic software follow-up, it is necessary to interrupt and re-mark it from the current moment, and then continue tracking until the marker disappears from the video. After all of the videos are tracked, the results are saved to obtain the position of each non-motorized vehicle in each frame relative to the initial coordinate origin, as shown in
Figure 4.
Step four: Video correction. To stabilize the video shoot, the traffic survey is conducted at a time when the weather is fine and there is no wind. Moreover, DJI drones boast strong hover-shooting stability and satisfactory image anti-shake effect, but the video still has slight jitter at a certain angle. To obtain more accurate data, the initial coordinates of each frame obtained in step three are projected along the non-motorized vehicle lane line and the vertical direction, and the XOY coordinate system is constructed, with the X axis along the non-motorized vehicle lane direction and the Y axis in the vertical direction. In the video tracking stage, three fixed points in the image are selected, as shown in the figure. The coordinates are respectively calibrated as
, and then the new coordinates
of non-motorized vehicles at the current moment can be obtained via coordinate conversion, as shown in Formula (1).
Step five: Data smoothing. To reduce the random fluctuation of space–time coordinates during vehicle tracking, the five-point difference method is used to smooth the coordinates of non-motorized vehicles, as shown in Formula (2). For example, after the coordinate information of non-motorized vehicles is obtained, the first derivative of the coordinates will give the speed of each non-motorized vehicle at every moment, as shown in Formula (3).
where:
—longitudinal position of a bicycle i at time t;
—lateral position of a bicycle i at time t;
—longitudinal position of a bicycle i at time t;
—lateral position of a bicycle i at time t.
After the above extraction process, the coordinates and speed information of each vehicle can be obtained. For the extraction of microscopic individual indicators, the manual information identification and recording method of professionals is adopted. In this study, eight graduate students were assigned to perform this work. Prior to information discrimination, information identification training should be carried out for recorders, and unified standards should be established in identifying different vehicle types, judging individual attributes of riders, and how distinguishing information quickly and accurately. Each road segment is equipped with two graduate students to ensure accurate and efficient data recording. By processing the data of 12 road segments separately, the following indicators can be obtained statistically, which lays the foundation for the follow-up research.
Speed: The distance traveled by vehicles in a unit of time. The average speed of the whole road segment is the arithmetic average of the speeds of all types of non-motorized vehicles on the non-motorized vehicle lane, which can reflect the average state of non-motorized vehicles on this road segment.
Interaction events: The overtaking/car-following interaction is observed between two non-motorized vehicles according to the video. According to previous studies, when the longitudinal distance between two vehicles involved in an interactive event is less than 5 m, it is recorded as an interactive event, and it is determined whether this interactive behavior is overtaking or following [
25]. The individual attributes of the overtaking vehicle and the overtaken vehicle should be identified from the camera.
Individual attributes of riders: judge the individual attributes of riders, gender (male/female), age (young/middle-aged/elderly), category of non-motorized vehicles (as shown in
Figure 5), human-carrying/large object-carrying, and professional delivery personnel such as courier/takeaway rider.
2.4. Data Analysis
In this study, the rider’s decision in overtaking behavior is taken as the dependent variable, and its values are binary, Y = 0, which indicates that the non-motorized vehicle rider chooses to follow the car, and Y = 1 indicates that the non-motorized vehicle rider chooses to overtake. The binary logit model shows strong adaptability in dealing with the situation that such dependent variables are binary variables.
A binary logit regression model was estimated to model the overtaking decision of non-motorized vehicle riders:
where
Xkn is the vector of explanatory variables,
βk is the vector of corresponding coefficients for explanatory variables, and
εkn is the identically and independently distributed random error term.
At the same time, the influence parameter coefficient was assumed to be random distribution, considering that the heterogeneity of the variables was not observed [
32,
33].
where
is the mean effect of the variable, and
represents normal distribution with the means of zero and variances of
.
Therefore, the overtaking probability of the rider’s random parameter logit model is constructed as:
To determine whether a variable could be selected as a random parameter, a stepwise iterative method was established [
34]. Every variable was evaluated in the model as a fixed or a random parameter. A statistical test of improved likelihood was used to determine the fitness. These processes continued until the model was stable to the change, and the model fit best.
In the present study, the variables ‘Age’ and ‘Whether it is a professional delivery worker’ were selected as random parameters. Hence, normal/lognormal/uniform distributional forms were tested, and the simulated maximum likelihood with 200 Halton draws was used to make the coefficient estimation computationally efficient. As a result, the normal distribution provided the best estimation results. The Akaike information criterion (AIC) was used to evaluate the fitting performance of the candidate models. The NLOGIT 5.0 statistical software (Econometric Software Inc., Greene, 2012) was used in the analysis.