Search

EP-4207131-B1 - AUTOMATED CUT-IN IDENTIFICATION AND CLASSIFICATION

EP4207131B1EP 4207131 B1EP4207131 B1EP 4207131B1EP-4207131-B1

Inventors

  • ABAD, Pablo
  • KALIA, AMAN VED
  • GOMEZ-RAMOS, JUAN
  • CHEN, JIAYI
  • MALIK, JASMAN SINGH
  • RIGGS, JOE

Dates

Publication Date
20260513
Application Date
20221221

Claims (11)

  1. A computer-implemented method for cut-in detection comprising: obtaining operational data about one or more vehicles, wherein the operational data about a given vehicle (500) of the one or more vehicles comprises information about sizes and locations over time of one or more additional vehicles (510) in an environment of the given vehicle, information about a location and a trajectory of the given vehicle over time, and information about an extent of a lane (505) being navigated by the given vehicle over time; based on the operational data, identifying the presence of one or more cut-ins within the operational data, wherein identifying the presence of a given cut-in comprises: (i) determining a location and an extent over time of a bounding box surrounding a particular additional vehicle in the environment of the given vehicle, (ii) determining that the particular additional vehicle entered the lane being navigated by the given vehicle by at least one of determining that at least one vertex of the bounding box was located more than a threshold distance within the lane being navigated by the given vehicle or determining that the bounding box overlapped with the lane being navigated by the given vehicle by more than a threshold amount, and (iii) determining that the ability of the given vehicle to maintain its course and speed was impeded by the particular additional vehicle entering the lane being navigated by the given vehicle; extracting, from the operational data, cut-in data that depicts one or more of the cut-ins identified within the operational data; and based on the extracted cut-in data, training a model for controlling an autonomous vehicle (100), wherein training the model for controlling the autonomous vehicle based on the extracted cut-in data comprises: determining one or more scores for the given vehicle's reaction to one or more of the cut-ins identified within the operational data, wherein the given vehicle's reaction to one or more of the cut-ins identified within the operational data was determined where a version of the model was controlling the given vehicle during the identified one or more cut-ins; and based on the one or more scores, updating the model.
  2. The method of claim 1, further comprising, based on the extracted cut-in data, classifying one or more of the cut-ins identified within the operational data.
  3. The method of claim 2, wherein classifying one or more of the cut-ins identified within the operational data comprises classifying one or more of the cut-ins identified within the operational data as at least one of a nominal cut-in of a slower vehicle, a nominal cut-in of a faster vehicle, an aborted cut-in, a cut-through, or a zipper cut-in.
  4. The method of claim 2, wherein classifying one or more of the cut-ins identified within the operational data comprises classifying one or more of the cut-ins identified within the operational data based on at least one of: a degree to which a vehicle slowed down due to the cut-in, a degree to which a vehicle applied brakes due to the cut-in, a timing of an action taken by a vehicle in reaction to the cut-in relative to a timing at which the cut-in occurred, a minimum distance between a vehicle and a cut-in agent following a cut-in, or a duration of a period that a vehicle followed a cut-in agent at a distance less than a threshold safety distance following a cut-in.
  5. The method of any preceding claim, further comprising, based on the one or more scores, generating human-interpretable metrics for the quality of the model with respect to safety, efficiency, or vehicle wear, wherein the model is one of a plurality of such models, and wherein updating the model comprises using the one or more scores to select whether to retain the model, replicate the model, fold the model, discard the model, combine the model with another model in the plurality of models, or set a weight of the model relative to other models in the plurality of models.
  6. The method of any preceding claim, wherein determining the one or more scores comprises generating a score based on at least one of: a degree to which the autonomous vehicle slowed down due to the particular cut-in, a degree to which the autonomous vehicle applied brakes due to the particular cut-in, a timing of an action taken by the autonomous vehicle in reaction to the particular cut-in relative to a timing at which the particular cut-in occurred, a minimum distance between the autonomous vehicle and a cut-in agent following the particular cut-in, or a duration of a period that the autonomous vehicle followed a cut-in agent at a distance less than a threshold safety distance following the particular cut-in.
  7. The method of any preceding claim, further comprising: storing the trained model in a computer readable medium of the autonomous vehicle.
  8. The method of any preceding claim, further comprising: controlling the autonomous vehicle using the trained model.
  9. A non-transitory computer readable medium (114) storing instructions (115) that, when executed by one or more processors (113) of a computing device, cause the computing device to perform operations comprising the method of any of claims 1-8.
  10. A system comprising: one or more processing units (113); and a non-transitory computer-readable medium (114) storing at least computer-executable instructions (115) that, when executed by the one or more processing units, cause the computing device to perform operations comprising the method of any of claims 1-8.
  11. An autonomous vehicle (100) comprising: a computer readable storage medium (114) storing a model for controlling an autonomous vehicle, wherein the model is trained according to the method of any of claims 1-6.

Description

BACKGROUND A controller for controlling an autonomous vehicle can be generated or improved based on training data depicting the real or simulated operation of a vehicle in a target environment (e.g., driving on a highway). Such training data can include sensor data (e.g., lidar data, GPS data, radar data, vehicle status data), commands to the vehicle (generated by an autonomous controller and/or by a human driver), or other data related to the operation of a vehicle (e.g., an autonomous vehicle) in a target environment and/or performing target behaviors. For instance, a method for cut-in event prediction and risk assessment for autonomous driving is known from Zhang Jinwei ET AL: "Interaction-Aware Cut-in Behavior Prediction and Risk Assessment for Autonomous Driving",IFAC-PapersOnLine - 16th IFAC Symposium on Control in Transportation Systems CTS 2021 Lille, France, 8-10 June 2021, vol. 53, no. 5, 1 January 2020 (2020-01-01), pages 656-663. SUMMARY In an aspect, a computer-implemented method for cut-in detection, as set out in appended claim 1, is provided. In another aspect, a non-transitory computer readable medium is provided as set out in appended claim 9. In another aspect, a system is provided as set out in appended claim 10. In a further aspect, an autonomous vehicle is provided according to claim 11. These as well as other aspects, advantages, and alternatives will become apparent to those of ordinary skill in the art by reading the following detailed description, with reference, where appropriate, to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 is a functional block diagram illustrating a vehicle, according to example embodiments.Figure 2A is an illustration of a physical configuration of a vehicle, according to example embodiments.Figure 2B is an illustration of a physical configuration of a vehicle, according to example embodiments.Figure 2C is an illustration of a physical configuration of a vehicle, according to example embodiments.Figure 2D is an illustration of a physical configuration of a vehicle, according to example embodiments.Figure 2E is an illustration of a physical configuration of a vehicle, according to example embodiments.Figure 3 is a conceptual illustration of wireless communication between various computing systems related to an autonomous vehicle, according to example embodiments.Figure 4A is a block diagram of a system including a lidar device, according to example embodiments.Figure 4B is a block diagram of a lidar device, according to example embodiments.Figure 5A is an illustration of vehicles navigating an environment, according to an example embodiment.Figure 5B is an illustration of an analysis of the vehicles and environment depicted in Figure 5A, according to an example embodiment.Figure 6A is an illustration of vehicles navigating an environment, according to an example embodiment.Figure 6B is an illustration of vehicles navigating an environment, according to an example embodiment.Figure 6C is an illustration of vehicles navigating an environment, according to an example embodiment.Figure 6D is an illustration of vehicles navigating an environment, according to an example embodiment.Figure 7 is a flowchart of a method, according to example embodiments.Figure 8 is a flowchart of a method, according to example embodiments.Figure 9 is a flowchart of a method, according to example embodiments. DETAILED DESCRIPTION Training data depicting the real or simulated operation of a vehicle in a target environment (e.g., a semi-trailer truck hauling a trailer, a passenger automobile, or some other vehicle driving on a highway or in some other environment) can be used to train or otherwise generate a model for controlling an autonomous vehicle. A very large amount of such training data is often available. However, it can be difficult to apply such data to the generation, updating, improvement, and/or training of a model for controlling an autonomous or semi-autonomous vehicle. This can be due in part to the computational cost of analyzing the entirety of such a training dataset. This difficulty can also be related to much of the training data representing "easy" driving conditions, e.g., steady driving on clear roads without interference from other vehicles. It can be desirable to emphasize training data that depicts more "difficult" circumstances, as these circumstances are often the most difficult to adequately address by a model for controlling an autonomous or semi-autonomous vehicle. One particular variety of such "difficult" driving conditions or events is the occurrence of a "cut-in." A cut-in is an event wherein another vehicle (a "cut-in agent") enters the lane ahead of (or behind) a target vehicle, interfering with the target vehicle's ability to maintain its course and speed. A cut-in differs from a "merge" event, in which another vehicle enters the lane ahead of (or behind) a target vehicle, in that such a merge event does not interfere with