Development of test scenarios and bicyclist surrogate for the evaluation of bicyclist automatic emergency braking systems

3.1 PP:VOT, parallel crashes with the vehicle traveling straight and overtaking the bicycle

The scenarios with the bicycle being overtaken by vehicle represent the crashes that do not happen at intersections. Consequently, the vehicle speed could be represented by a function related to the speed limit of the road. The term BCType of GES/FARS provides details for decomposing the PP:VOT into subtypes by incorporating bicycle behavior beyond the direction of travel (Figure 3). The subtypes include the bicycle turning right in front of the vehicle (PP:VOT-BRT), the bicycle turning left in front of the vehicle (PP:VOT-BLT), the bicycle traveling straight and struck from behind by the vehicle (PP:VOT-BS) and the bicycle riding out from a midblock location (PP:VOT-BRO).

Table II shows the data of crash geometries PP:VOT. The overall social cost in this group is US$1812.22m, covering about 26 per cent of the total annual social cost. In the sub-scenarios, PP:VOT-BS comprises the maximum percentage of the vehicle overtaking crashes; it covers approximately 19 per cent of the total social cost of all bike crashes. It should be noticed that “Fatalities” in Table II is annualized fatalities, and the “per cent” is the percentage of social cost in all crash cases. The unit for social cost is million US dollars.

3.2 PP:VHO, vehicle going straight and has head-on collisions with the bicycle

Head-on collisions are typically not associated with intersections and both bicyclist and vehicle travel at reasonably high speeds (Figure 4). Details for head-on crashes include:

• Bicyclist is making a right turn (PP:VHO-BRT), where it is initially on the opposite side of the road.

• Bicyclist is initially traveling on the same side of the road as the vehicle but does not leave sufficient gap for the vehicle (PP:VHO-BLT).

• Vehicle is traveling on the wrong side of the road, or the bicycle is traveling on the wrong side of the road (PP:VHO-BS).

Since both mistakes involve the same crash geometries, they are combined into one single scenario. The results are shown in Table III. It covers about 4.3 per cent of the annual social cost.

3.3 CP:VDO, crossing paths with vehicle going straight without the right of way

This type of crashes represents scenarios that the vehicle is traveling straight and does not have the right of way (Figure 5). The detailed sub-scenario types are shown in Table IV. This group represents about 13 per cent of the bicycle crashes and less than 1 per cent of the fatalities. Two different sub-scenario types are found: 1) CP:VDO-BS (driver stops and goes and fails to yield) and CP:VDT-BS (driver does not stop), which represent 3.9 and 1.2 per cent of the total social cost, respectively.

3.4 CP:VS, crossing paths with vehicle traveling straight with right of the way

This type of crashes represents the scenarios where the vehicle is traveling straight and has the right of way (Figure 6). It is expected that the vehicle travels within the speed limit. This type of crashes is very complicated. Except for the sub-scenarios of bicycle riding through (CP:VS-BDT) and riding out (CP:VS-BRO), the crashes involve the bicyclist cutting the corner (CP:VS-BCC), the bicyclist swinging wide (CP:VS-BSW), the bicyclist is trapped or there is an obscuring object (CP:VS-BMT) and other cases where the bicyclist has turning errors (CP:VS-BTE). The results are shown in Table V. We can find that the scenario of CP:VS-BRO has the highest percentage in terms of the social cost, which covers about 12.5 per cent of the total social cost.

3.5 PP:VLT, parallel scenarios with the vehicle turning left

This type of crashes is shown in Figure 7, which has two common sub-scenario types:

1. vehicle turning left and the bicyclist traveling in the same direction as the vehicle before it turns (PP:VLT-BS); and

2. bicyclist on the road or on the sidewalk enters the intersection, while the vehicle turns left into the roadway or a junction (PP:VLT-BHO).

These two types of crashes comprise about 8 per cent of all bicyclist crashes and 4.8 per cent of the total social cost (Table VI).

3.6 CP:VLT, crossing scenarios with vehicle turning left

The second type of vehicle left-turn scenarios is associated with bicycle crossing paths (Figure 8). This type of crashes includes:

• vehicle does not yield to a bicycle (CP:VDL-BS);

• driver cuts the corner and strikes a bicyclist (CP:VCC-BS);

• bicycle drives out (CP:VLT-BDO); and

• other unidentified cases.

As this type of crashes only covers about 1.9 per cent of the total social cost, they are not considered in the AEB system evaluation (Table VII).

3.7 CP:VRT, crossing scenarios with vehicle turning right

This type of crashes occurs when the vehicle turns right and bicyclist crosses (Figure 9). It covers about 8 per cent of the total social cost. The most frequent scenario in this type is the situation where the vehicle swings too wide as they make the right turn (CP:VSW-BS). It covers 3.1 per cent of the total social cost. The second most frequent crash scenario is CP:VDR-BS, where the vehicle “drives out” while making a right turn and bicyclist goes straight and crosses the road (Table VIII).

3.8 PP:VRT, parallel scenarios with the vehicle turning right

This type of crashes involves parallel path crashes with the vehicle turning right. They consist of two bicyclist crash scenarios: PP:VRT-BS and PP:VRT-BHO. As both of them have a small percentage of the total social cost (2.2 and 1.0 per cent, respectively), they are not considered in the AEB system evaluation (Figure 10) (Table IX).

3.9 Recommended crash geometries

Based on the detailed analysis of the above scenarios, a set of five crash scenarios is recommended for the evaluation of the bicyclist AEB systems. It involves three parallel crash geometries:

1. parallel paths with the vehicle overtaking the bicycle (PP:VOT);

2. parallel paths with the vehicle approaching head-on (PP:VHO); and

3. parallel paths with vehicle turning left and bicycle striking vehicle head-on (PP:VLT-BHO).

It also involves two crossing crash geometries:

1. crossing paths with vehicle driving out (CP:VDO-BS); and

2. crossing paths with vehicle going straight and the bicyclist is riding out (CP:VS-BRO).

While the CP:VS-BDT scenario incorporates a large fraction of overall social cost, it is very similar to the CP:VS-BDT scenario; the only difference being whether the bicycle stops or not prior to its failure to yield at the intersection. This allows the inclusion of the CP:VDT-BS scenario where the vehicle is primarily at fault by failing to yield.

4.1 Lighting conditions

Three lighting levels are considered: daylight, dark-lit (which includes dusk and dawn) and dark-unlit. The statistical data show that a vast majority of bicyclist crashes occur during the daytime, which cover 61.7 per cent of the total social cost (Table X).

4.2 Obscuring objects

Our analysis of obscuring objects considers three major types:

1. physical objects between the driver and the bicyclist, such as a parked car or truck;

2. moving vehicles during a multiple threat/trapped event, which mostly happen in crossing pathway geometries; and

3. glare.

The crashes involving obscuring objects account for less than 10 per cent of all crashes (Table XI). The detailed scenarios suggest that obscuring objects do not play an important role except for the scenario of CP:VS-BMT.

4.3 Vehicle speed

There is no data source available to describe vehicle speed for bicyclist-related crashes. We estimate the vehicle crash speed by using the speed limit of the road. The distribution of the speed limit for crashes is shown in Figure 11, where we find that the most important speed limit for crashes is 25 mph, while the most important speed limits for fatalities are 35 mph and 45 mph.

4.4 Bicyclist speed

From the crash databases, it is difficult to estimate bicycle speed during crashes. TASI 110-car NDD were used for finding bicyclist speed (Fu et al., 2017). From the data, 1,000 bicyclist cases are obtained through video analysis. Three main scenarios are analyzed:

1. bicyclist moving along the road;

2. bicyclist crossing the road with constant speed (ride through); and

3. bicyclist crossing the road from stationary (ride out).

The bicycle moving speed distributions for the scenarios of moving along the road and ride through crossing the road are shown in Figures 12 and 13, respectively. For the scenario of moving along the road, the average bicyclist moving speed is 5.59 m/s. The 25th and 75th percentile speeds are 4.06 m/s and 6.94 m/s, respectively. For the scenario of ride through crossing the road, the average moving speed is 5.23 m/s, and the 25th and 75th percentile speeds are 3.94 m/s and 6.26 m/s, respectively. The scenario of crossing the road from being stationary is more complex. The overall average moving speed is 3.5 m/s. To obtain detailed moving behavior, the road is marked with five key points: roadside, 25 per cent of the road, 50 per cent of the road, 75 per cent of the road and another side of the road. The average speeds when crossing the road from stationary at 25, 50 and 75 per cent of the road are 2.95 m/s, 3.77 m/s and 3.85 m/s, respectively. The average bicyclist crossing the road acceleration is 1.4 m/s².

5.1 Representative size of bicycle riders and bicycles

According to “Bicycling and Walking in the USA, 2014 Benchmarking Report”, adults between the age of 16 and 64 account for 77 and 68 per cent of all bicyclist fatalities and injuries, respectively (Milne and Melin, 2014). Seniors of age 65 or higher represent 12 per cent of all bicyclist fatalities and 5 per cent of all bicyclist injuries. Children under age of 16 represent 11 per cent of all bicyclist fatalities and 27 per cent of all bicyclist injuries between 2009 and 2011. As bicyclists over the age of 16 cover 89 per cent of fatalities and 73 per cent of injuries, only the adult surrogate bicyclist size is recommended. Moreover, only the male adult-sized bicyclist surrogate is recommended, as male account for 87 per cent of bicyclist fatalities and 83 per cent of bicyclist injuries.

For scientific soundness, a more precise method to determine the size of the surrogate is the weighted 50th percentile size of male and female adults. Therefore, the weighted height of US male and female adults, 173 cm, is recommended as the height of bicycle rider surrogate. As a few centimeters height difference might not affect the bicyclist detection significantly, traditional 50th percentile height of US male adults, 175.6 cm (CDC, 2018), is also acceptable as the height of surrogate bicycle rider.

According to the statistical report of Bicycle Product Supplier Association in 2012, the most popular adult bicycles in US market have 26-inch wheel size. The frame types of most adult bicycles are mountain bicycles and road bicycles. Therefore, the wheel size of the adult surrogate bicycle is suggested to be 26 inches, and the shape of the adult surrogate bicycle is defined as a mix of mountain and road types.

5.2 Limb motion of the bicycle rider

Limb motion of bicyclists can be a useful feature for bicyclist detection. Many camera-based pedestrian and bicyclist detection studies emphasize the importance of limb motion for achieving better detection results (Wojek et al., 2009; Takahashi et al., 2010). For the radar-based detection systems, it was also pointed out that the limb motion/pedaling can produce a very noticeable micro-Doppler effect from the front (0^°), back (180^°) and 45^°-side observation angles, as shown in Figure 14 (Belgiovane and Chen, 2017).

There has been some discussion about whether the limb motion/pedaling is required for the surrogate bicyclist. Euro NCAP has adopted the bicyclist surrogate with fixed leg posture (both legs bent) based on their study that suggests that a majority of bicyclists (over 80 per cent) stop pedaling when crossing an intersection in Europe. By examining 484 randomly selected bicyclist video clips in the TASI 110-car NDD (Sherony et al., 2016), it is observed that 83.2 per cent of bicyclists have pedaling motion when crossing the road, and 100 per cent of bicyclists have pedaling motion when moving along the road. The average pedaling frequency is 0.85 Hz for the cases of crossing the road (CP:VDO and CP:VS-BRO) and 1.01 Hz for the cases of moving along the road (PP-VOT and PP-VHO).

To demonstrate the importance of pedaling motion in the performance of bicyclist AEB system, a comparison AEB test of the same surrogate bicyclist with two different leg postures (one straight leg or both bending legs) was conducted [Figure 15(a) and Figure 15(b)]. The test vehicle used was a commercially available SUV that has a radar and camera-based bicyclist AEB system. The results [plotted in Figure 15(c)] show that the AEB system performs better for the cases where one leg is straight than for cases where both legs are bent. The results demonstrate that leg postures do affect the performance of bicyclist AEB systems. The same conclusion was also reported by a study from Euro NCAP (den Camp et al., 2017). The possible reason of performance difference given in den Camp et al. (2017) is that the AEB system uses characteristics from the 360^° pedaling sequence. Based on the above discussions, the leg motion/pedaling capability is recommended for the proposed surrogate bicyclist design.

5.3 Wheel rotation

Although it is difficult to know if any camera-based AEB system on a production vehicle uses the wheel rotation to detect the bicyclist, it is well documented that radar systems use micro-Doppler to detect wheel rotation of bicyclists. The pedaling motion and the wheel rotation can generate clear micro-Doppler responses. The experimental results presented in Belgiovane and Chen (2017) have shown that a rotating wheel could produce distinctive periodic micro-Doppler spectral lines whose fundamental frequency is related to the vehicle speed, and the periodic frequency changes are related to the leg and wheel rotations. This micro-Doppler effect of a wheel can be observed in front (0°), back (180°) and 45° side view angles. Figure 16 shows the micro-Doppler measurement of rotating bicycle wheel of 66 rpm in the back view. The discrete ripples in the vertical direction reflect the tire thread pattern.

When both wheels can be detected using micro-Doppler, an H-shape micro-Doppler signature can be observed (CDN, 2018; TNO, 2018). This H-shaped bicycle signature cannot be observed when the bicycle is ahead of the vehicle and moving along the road.

As the automotive radar systems can use micro-Doppler characteristics for bicycle detection, the wheel rotation is also recommended for the proposed surrogate bicyclist design.

5.4 Clothing color of bicycle rider

The camera is a common sensor used for bicyclist detection. Besides the shape of the bicyclist, the clothing color also significantly affects the performance of bicyclist AEB systems. To ensure that the clothing color identification is not affected by lighting conditions, the images of 1,905 adult bicyclists not under shade were filtered from all bicyclists detected in the TASI 110 car naturalistic driving study. For these bicyclists’ images, the K-means clustering algorithm is applied to find the color clusters for both upper and lower cloth colors in the International Commission on Illumination (CIE) LUV color space. As a result, the black/deep gray combination is suggested to be the most representative clothing color for adult bicyclists. Details of color selection method can be found in den Camp et al. (2017). For the convenience of selecting the color needed, a range of ±10 per cent variation of brightness is used to make the acceptable color range. The RGB values and acceptable color ranges for both upper and lower clothing color are shown in Figure 17.

5.5 Surrogate bicyclist hardware development

The surrogate bicyclist (including a bicycle rider and a bicycle surrogate) is designed to represent the visual and radar characteristics of the real bicyclists in the USA. Therefore, the surrogate bicyclist should have similar physical properties with respect to most sensors used for the bicyclist detection. The parameters identified and recommended in Sections 5.1-5.4 are considered as part of the requirements for the proposed surrogate bicyclist design.

In 2015, the first generation of the surrogate bicyclist was developed by TASI (Yi et al., 2016). The surrogate bicycle rider has skin that matches the 77- to 78-GHz radar reflectivity of the human skin. By using this skin and realistic body shape, the radar cross section (RCS) of the surrogate bicycle rider is similar to representative real bicycle rider from all 360-degree angles in the view of the 77 GHz automotive radar. To harmonize our surrogate bicyclist design with CATS’ design, we conducted a comparison study by cross-testing and examining each other’s surrogate bicyclist prototype in 2016. Based on the results, we modified our design and developed a new generation of the harmonized surrogate bicyclist, which consists of three key components: bicycle rider, bicycle and transport sled (Figure 18). Owing to the fact that most adult bicyclists involved in crashes in the USA are male, the 50th percentile height of US male population, 175.6 cm, is considered as the surrogate bike rider’s height. The mannequin for pedestrian PCS evaluation (Yi et al., 2014) is modified as the bicycle rider. The body and limbs are made to generate realistic human body shape. The harmonized surrogate bicycle has both leg pedaling and wheel rotation to produce the proper micro-Doppler features (Belgiovane and Chen, 2017) and support the camera-based AEB systems (Takahashi et al., 2010).

The finished prototype of the harmonized surrogate bicyclist is shown in Figure 18. The total weight of the surrogate bicycle is 18.4 lbs (8.4 kg), and the total weight of the surrogate bicycle rider is 9 lbs (4.1 kg). The crash testing experiments show that a developed target can handle 45-mph full-speed crash without damage or with minor damage and can be reset in 5 min.

5.6 Radar Cross Section and micro-Doppler effect

The far-field RCS pattern data of the harmonized surrogate bicyclist are measured. Figure 19(a) shows the comparison between the smoothed (7^° average moving) RCS patterns of the US 26-inch mountain bike with a human rider and the prototype harmonized surrogate bicycle and rider mannequin. Figure 19(b) compares the measured RCS patterns of the 26-inch mountain bike and the harmonized surrogate bicycle. It can be seen that the measured RCS data of the developed surrogate bicyclist agree well with that of the real rider plus bicycle in terms of both RCS pattern shape and level in 360 degrees.

Figures 20 and 21 show the micro-Doppler measurement of leg pedaling and wheel rotation produced by the developed surrogate bicyclist. The micro-Doppler features match the real human leg pedaling shown in Figure 14, and the real wheel rotation is shown in Figure 16.