Methods and systems for handling occlusions in operation of autonomous vehicle
An autonomous vehicle navigates an environment in which occlusions block the vehicle's ability to detect moving objects. The vehicle handles this by receiving sensor data corresponding to the environment, identifying one or more particles of a plurality of initialized particles that have a non-zero probability of being associated with a potential occluded actor in the occluded region, and generating a trajectory of the autonomous vehicle for traversing the environment based on spatiotemporal reasoning about the occluded region and taking into account the one or more particles. Each particle may be associated with a potential actor.
1 . A method of navigating an autonomous vehicle through an environment comprising an occluded region, the method comprising: generating a plurality of initialized particles by: initializing, using sampling, a plurality of particles in a region of interest using map data and absent sensor data, each particle being representative of an initial hypothesis associated with a potential actor; and assigning, to each of the plurality of particles, an initial hypothesis state and an initial probability using a uniform distribution sampling and absent the sensor data; receiving the sensor data corresponding to the environment; generating, using the sensor data, a visibility map corresponding to the environment, the visibility map comprising a plurality of cells that each include information about an occluded state of that cell and being representative of an ability of the autonomous vehicle to perceive or track one or more objects in the environment; generating, for each of the plurality of initialized particles, a forecasted hypothesis state by propagating the respective initialized particle using a process model associated with the particle to a measurement time associated with the sensor data; assigning, using the visibility map and the forecasted hypothesis state, an updated probability to each of the plurality of initialized particles; identifying, based on the updated probabilities, one or more particles among the plurality of initialized particles having a non-zero probability as being associated with a potential occluded actor in the occluded region, each particle being associated with a potential actor; and generating, based on spatiotemporal reasoning about the occluded region and taking into account the one or more particles, a trajectory of the autonomous vehicle for traversing the environment; and controlling the autonomous vehicle to traverse the environment using the trajectory generated.
2 . The method of claim 1 , wherein the region of interest comprises a conflict region that intersects with a lane of travel of the autonomous vehicle through the environment.
3 . The method of claim 2 , wherein the initial hypothesis state of each particle comprises at least one of the following: a position, a velocity, an acceleration, or an actor type of the potential actor.
4 . The method of claim 2 , wherein the potential actor is selected from at least one of the following: a pedestrian, a bicyclist, or a vehicle.
5 . The method of claim 1 , further comprising:
identifying a second one or more of the plurality of initialized particles as particles that have a zero updated probability as being associated with potential visible tracked actors; and
discarding the second one or more of the plurality of initialized particles.
6 . The method of claim 1 , further comprising generating, based on the one or more particles, a phantom actor corresponding to a worst-case potentially occluded actor in the occluded region.
7 . The method of claim 1 , further comprising generating, based on the one or more particles, an unknown region frontier that causes the autonomous vehicle to come to a stop before the unknown region frontier.
8 . A system for navigating an autonomous vehicle, the system comprising: at least one processor; and programming instructions stored in a memory and configured to cause the processor to: generate a plurality of initialized particles by: initializing, using sampling, a plurality of particles in a region of interest using map data and absent sensor data, each particle being representative of an initial hypothesis associated with a potential actor; and assigning, to each of the plurality of particles, an initial hypothesis state and an initial probability using a uniform distribution sampling and absent the sensor data; receive the sensor data corresponding to an environment; generate, based on the sensor data, a visibility map corresponding to the environment, the visibility map comprising a plurality of cells that each include information about an occluded state of that cell and being representative of an ability of the autonomous vehicle to perceive or track one or more objects in the environment; generate, for each of the plurality of initialized particles, a forecasted hypothesis state by propagating the respective initialized particle using a process model associated with the particle to a measurement time associated with the sensor data; assign, using the visibility map and the forecasted hypothesis state, an updated probability to each of the plurality of initialized particles; identify, based on the updated probabilities, one or more particles among the plurality of initialized particles having a non-zero probability as being associated with a potential occluded actor in an occluded region of the environment, each particle being associated with a potential actor; generate, based on spatiotemporal reasoning about the occluded region and taking into account the one or more particles, a trajectory of the autonomous vehicle for traversing the environment; and control the autonomous vehicle to traverse the environment using the trajectory generated.
9 . The system of claim 8 , wherein the region of interest comprises a conflict region that intersects with a lane of travel of the autonomous vehicle through the environment.
10 . The system of claim 8 , wherein the initial hypothesis state of each particle comprises at least one of the following: a position, a velocity, an acceleration, or an actor type of the potential actor.
11 . The system of claim 8 , wherein the potential actor is selected from at least one of the following: a pedestrian, a bicyclist, or a vehicle.
12 . The system of claim 8 , further comprising additional programming instructions that are configured to cause the processor to:
identify a second one or more of the plurality of initialized particles as particles that have a zero updated probability as being associated with potential visible tracked actors; and
discard the second one or more of the plurality of initialized particles.
13 . The system of claim 8 , further comprising additional programming instructions that are configured to cause the processor to generate, based on the one or more particles, a phantom actor corresponding to a worst-case potentially occluded actor in the occluded region.
14 . A computer program product comprising a non-transitory computer-readable medium that stores instructions that, when executed by a computing device, will cause the computing device to perform operations comprising: generating a plurality of initialized particles by: initializing, using sampling, a plurality of particles in a region of interest using map data and absent sensor data, each particle being representative of an initial hypothesis associated with a potential actor; and assigning, to each of the plurality of particles, an initial hypothesis state and an initial probability using a uniform distribution sampling and absent the sensor data; receiving the sensor data corresponding to an environment of an autonomous vehicle; generating, using the sensor data, a visibility map corresponding to the environment, the visibility map comprising a plurality of cells that each include information about an occluded state of that cell and being representative of an ability of the autonomous vehicle to perceive or track one or more objects in the environment;
generating, for each of the plurality of initialized particles, a forecasted hypothesis state by propagating the respective initialized particle using a process model associated with the particle to a measurement time associated with the sensor data; assigning, using the visibility map and the forecasted hypothesis state, an updated probability to each of the plurality of initialized particles; identifying, based on the updated probabilities, one or more particles from among the plurality of initialized particles having a non-zero probability as being associated with a potential occluded actor in an occluded region of the environment, each particle being associated with a potential actor; generating, based on spatiotemporal reasoning about the occluded region and taking into account the one or more particles, a trajectory of the autonomous vehicle for traversing the environment; and controlling the autonomous vehicle to traverse the environment using the trajectory generated.