WATonoBus: Field-Tested All-Weather Autonomous Shuttle Technology

TL;DR

This work proposes a multi-module and modular system architecture with considerations for adverse weather across the perception level, through features such as snow covered curb detection, to decision-making and safety monitoring for all-weather autonomous vehicle operation.

Abstract

All-weather autonomous vehicle operation poses significant challenges, encompassing modules from perception and decision-making to path planning and control. The complexity arises from the need to address adverse weather conditions such as rain, snow, and fog across the autonomy stack. Conventional model-based single-module approaches often lack holistic integration with upstream or downstream tasks. We tackle this problem by proposing a multi-module and modular system architecture with considerations for adverse weather across the perception level, through features such as snow covered curb detection, to decision-making and safety monitoring. Through daily weekday service on the WATonoBus platform for almost two years, we demonstrate that our proposed approach is capable of addressing adverse weather conditions and provide valuable insights from edge cases observed during operation.

Figures (10)

• Figure 1: Illustration of WATonoBus sensor suite, compute system, control interface, and visualization utilities.
• Figure 2: A schematic of the overall software structure in WATonoBus.
• Figure 3: An example for geese detection. Points from the geese are correctly segmented as shown in red, and their corresponding convex hulls are shown in green polylines.
• Figure 4: (Upper left) Camera view. (Lower left) Point cloud view. Red points denote closer distance to ego. (Right) Road boundary detection results in heavy snow. Red points are detected curb points from LiDAR point cloud. Blue squares are estimated road boundary. Yellow line denotes curb position from HD map. The blue boxes are on the left of the yellow line, reflecting actual drivable space.
• Figure 5: A schematic of the overall late fusion algorithm. The radar provides longitudinal and lateral positions and velocities $x_r, y_r, v_x$, and $v_y$ of objects. After ground removal and clustering, we have object positions and the dimensions from LiDAR-only detection algorithm, $x_{Li}, y_{Li}, z_{Li}, w, h$, and $l$. From YOLOv4 and camera-based object detection we have bounding box dimensions and class $x_i, y_i, w_i, h_i$, $c$. The fusion process completes with frustum association and the final output is bounding box position, dimension, velocity and class $x,y,z,w,h,l,v_x,v_y$ for each object.