V2X-Communication-Aided Autonomous Driving: System Design and Experimental Validation

Abstract

In recent years, research concerning autonomous driving has gained momentum to enhance road safety and traffic efficiency. Relevant concepts are being applied to the fields of perception, planning, and control of automated vehicles to leverage the advantages offered by the vehicle-to-everything (V2X) communication technology. This paper presents a V2X communication-aided autonomous driving system for vehicles. It is comprised of three subsystems: beyond line-of-sight (BLOS) perception, extended planning, and control. Specifically, the BLOS perception subsystem facilitates unlimited LOS environmental perception through data fusion between local perception using on-board sensors and communication perception via V2X. In the extended planning subsystem, various algorithms are presented regarding the route, velocity, and behavior planning to reflect real-time traffic information obtained utilizing V2X communication. To verify the results, the proposed system was integrated into a full-scale vehicle that participated in the 2019 Hyundai Autonomous Vehicle Competition held in K-city with the V2X infrastructure. Using the proposed system, the authors demonstrated successful completion of all assigned real-life-based missions, including emergency braking caused by a jaywalker, detouring around a construction site ahead, complying with traffic signals, collision avoidance, and yielding the ego-lane for an emergency vehicle. The findings of this study demonstrated the possibility of several potential applications of V2X communication with regard to autonomous driving systems.

Keywords: V2X communication; autonomous driving system; control; intelligent transportation system; perception; planning.

Figure 1

Full-scale autonomous vehicle platform, Eurecar, in K-city.

Figure 2

Overall architecture of the Eurecar autonomous vehicle system. BLOS, beyond LOS.

Figure 3

Visualization of local perception result. (a) Vision-based multi-class detection result obtained using Drivenet. (b) Clustered LiDAR data in the 3D coordinate system using the DBSCAN algorithm. Using vision-based detection result, clusters corresponding to dynamic objects (pedestrian and neighboring vehicles) can be tracked for estimating relative velocity. (c) Cost map obtained from occupancy grid information with yellow indicating the grid to be occupied by obstacles. In general, the grid cost increases as the color changes from white to black.

Figure 4

Schematic of potential hazard calculation using dynamic obstacles.

Figure 5

Visualization of centerline-based road network (nodes and edges) in K-city with the route planning result. It was assumed that the global path planner received real-time traffic situation information (construction site locations) through V2I. The roadway centerline from the HD map is represented by the solid line. Dashed lines with arrows indicate the joint path and its direction. White circles indicate the nodes. The table lists the distance-level and traffic-level costs of each edge. By setting the traffic-level cost corresponding to the construction location to infinity, our global path planner planned the minimum-cost path to detour the section.

Figure 6

Road-model-based path candidates and global referential path in the Cartesian coordinate system.

Figure 7

FSM for behavior planning of Eurecar.

Figure 8

Schematic of longitudinal low-level control. Vehicle speed is controlled in a cascade manner with an outer speed loop and an inner pedal loop.

Figure 9

(a) Schematic of the kinematic model for the proposed path-following controller. (b) Experimental step response of original Stanley method with $L$ = 2.65 m with the gain k = 0.5/s and speed v = 6 m/s, showing cross-track error in meters and steering wheel angles in revolutions. (c) Experimental step response of the proposed method with $L$ = 2.65 m and $L_p$ = 5.3 m. Faster convergence with less overshoot was observed using the proposed method.

Figure 10

2019 Hyundai AVC route in K-city.

Figure 11

Result of Mission 1. At t1, Eurecar ran along the global path. At time t2 where the potential hazard exceeded a certain threshold, a deceleration command was sent to the low-level controller. The maximum potential hazard due to control delay and vehicle characteristics was at t3. At t4, the pedestrian stopped in the middle of the roadway. Eurecar stopped until the pedestrian crosses.

Figure 12

Result of Mission 2. Global locations of the construction sites received via V2I are marked with red dots, whereas blue and red lines indicate the initially provided path and the replanned path that made a detour around the construction sites, respectively.

Figure 13

Velocity profile and longitudinal control result in Mission 3. The top three rows depict the traffic light status at each intersection. Dotted and solid lines indicate remaining time received via V2I (1 Hz) and upsampled results (10 Hz) using an internal clock. At t1, the behavior planner changed the state to the intersection state. At t2, Eurecar drove through the first intersection without stopping. Because the remaining time of the red traffic light of the subsequent intersection was short, further deceleration was not required, and acceleration was performed in consideration of the signal state at the last intersection (t3). In the same way (from t1 to t2), the Eurecar passed the last intersection without stopping (from t4 to t5).

Figure 14

Results of collision avoidance mission. At time t1, Eurecar moved along the minimum-cost path among the candidates around the global path. At time t2, all candidate paths collided and were not feasible, thus leading to the creation of a collision avoidance path with the graph-based approach. From t3 to t5, Eurecar tracked the path generated by the graph-based approach, and at t6, Eurecar merged to the initial global path developed through the road-model-based approach.

Figure 15

Results of Mission 5. At time t1, Eurecar received the basic safety message (BSM) of the emergency vehicle approaching from behind while driving along the ego-lane. At times t2 and t3, the emergency vehicle approached within 10m, and a lane change to the right was performed by arbitrarily increasing the cost of path candidates generated around the ego-lane to yield lanes. At times t4 and t5, the emergency vehicle had not yet passed, and Eurecar ran along the changed lane. At time t6, the Eurecar changed the lane to the left to return to the original lane.