Hello Apollo development team. We used the CARLA simulator to model raw LiDAR data under foggy and rainy conditions.Based on this, we tested Apollo8.0 under different weather conditions. We discovered several potential bugs in the Apollo perception module under adverse conditions. Please see the details described below.
Describe the bug
Here are four major bug cases we identified:
Failure to Detect Small Obstacles: In adverse weather conditions, the AV was unable to detect pedestrians and cyclists.
Discontinuous Obstacle Detection: The AV experienced intermittent detection of obstacles in bad weather conditions.
Failure to Detect Large Vehicles: The AV could not detect large vehicles in adverse weather conditions.
Failure to Detect Obstacles Too Close: In adverse weather conditions, obstacles that were too close to the AV were only partially detected by the LiDAR, preventing proper recognition of these obstacles.
You can find in the example below:
In the figure below, under rainy conditions, a pedestrian or a cyclist is positioned on the path of the AV. The LiDAR's detection performance for small targets is substantially impaired, leading to the AV's failure to detect the pedestrian or cyclist. Consequently, this failure results in a collision. It's important to note that under clear weather conditions, this same situation wouldn't result in a collision.
In the figure below, the NPC vehicle is stationary at the intersection, while the AV plans to proceed straight through the intersection, with both vehicles occupying the same lane. Due to foggy conditions, the detection range of the AV's perception system is diminished. The AV only detects the NPC vehicle when it's relatively close, prompting the AV to decelerate and eventually stop. However, as the AV slows down, it continues to approach the NPC vehicle until the NPC vehicle enters the range of fog noise points surrounding the AV. Consequently, the AV fails to detect the NPC vehicle, leading to a collision.
In the figure below, an NPC truck is stationary in the right lane, while the AV plans to execute a right turn from the left lane. However, due to the adverse weather conditions causing reduced detection range and heightened detection noise in the LiDAR point cloud, the AV is unable to perceive the presence of the truck. Consequently, this leads to a collision between the AV and the stationary truck. Under clear weather conditions, such a collision would likely be avoided as the AV's perception system would function more effectively.
In the figure below, the AV encounters an NPC vehicle ahead, which is initially far away and moving slower than the AV. The AV intends to change lanes to the right before making a right turn at the intersection. Initially, the AV detects the NPC vehicle while it's still at a distance, so it plans to execute the lane change first and then proceed to follow the NPC vehicle for the right turn. However, as the AV executes the lane change maneuver, the distance between the AV and the NPC vehicle decreases. Consequently, the NPC vehicle enters the range of fog noise points surrounding the AV. At this critical point, the AV fails to detect the NPC vehicle, ultimately resulting in a collision.
Hello Apollo development team. We used the CARLA simulator to model raw LiDAR data under foggy and rainy conditions.Based on this, we tested Apollo8.0 under different weather conditions. We discovered several potential bugs in the Apollo perception module under adverse conditions. Please see the details described below.
Describe the bug Here are four major bug cases we identified: Failure to Detect Small Obstacles: In adverse weather conditions, the AV was unable to detect pedestrians and cyclists. Discontinuous Obstacle Detection: The AV experienced intermittent detection of obstacles in bad weather conditions. Failure to Detect Large Vehicles: The AV could not detect large vehicles in adverse weather conditions. Failure to Detect Obstacles Too Close: In adverse weather conditions, obstacles that were too close to the AV were only partially detected by the LiDAR, preventing proper recognition of these obstacles. You can find in the example below: In the figure below, under rainy conditions, a pedestrian or a cyclist is positioned on the path of the AV. The LiDAR's detection performance for small targets is substantially impaired, leading to the AV's failure to detect the pedestrian or cyclist. Consequently, this failure results in a collision. It's important to note that under clear weather conditions, this same situation wouldn't result in a collision.
In the figure below, the NPC vehicle is stationary at the intersection, while the AV plans to proceed straight through the intersection, with both vehicles occupying the same lane. Due to foggy conditions, the detection range of the AV's perception system is diminished. The AV only detects the NPC vehicle when it's relatively close, prompting the AV to decelerate and eventually stop. However, as the AV slows down, it continues to approach the NPC vehicle until the NPC vehicle enters the range of fog noise points surrounding the AV. Consequently, the AV fails to detect the NPC vehicle, leading to a collision.
In the figure below, an NPC truck is stationary in the right lane, while the AV plans to execute a right turn from the left lane. However, due to the adverse weather conditions causing reduced detection range and heightened detection noise in the LiDAR point cloud, the AV is unable to perceive the presence of the truck. Consequently, this leads to a collision between the AV and the stationary truck. Under clear weather conditions, such a collision would likely be avoided as the AV's perception system would function more effectively.
In the figure below, the AV encounters an NPC vehicle ahead, which is initially far away and moving slower than the AV. The AV intends to change lanes to the right before making a right turn at the intersection. Initially, the AV detects the NPC vehicle while it's still at a distance, so it plans to execute the lane change first and then proceed to follow the NPC vehicle for the right turn. However, as the AV executes the lane change maneuver, the distance between the AV and the NPC vehicle decreases. Consequently, the NPC vehicle enters the range of fog noise points surrounding the AV. At this critical point, the AV fails to detect the NPC vehicle, ultimately resulting in a collision.