Toyota Research Institute, Inc.

Toyota Jidosha Kabushiki Kaisha

Leland Stanford Junior University

Toyota Motor Engineering & Manufacturing North America, Inc.

Kuan-Hui Lee

Kun-Hsin Chen

Adrien David Gaidon

Arjun Bhargava

Chao Fang

Sudeep Pillai

Charles Christopher Ochoa

Jie Li

Jia-En M Pan

Vitor Guizilini

Karttikeya Mangalam

Rares A Ambrus

Shunsho Kaku

Juan Carlos Niebles Duque

Yusuke Kanzawa

German Ros Sanchez

Ehsan Adeli-Mosabbeb

Danil V Prokhorov

Michael R James

Xue Mei

Jeffrey Michael Walls

Naoki Nagasaka

Daniele Molinari

Hai Jin

Blake Warren Wulfe

Peiyan Gong

Steven A Parkison

Matthew Derry

Matthew T Kliemann

Boris Ivanovic

Logan Michael Ellis

Pavel V Tokmakov

Haofeng Chen

Takaaki Tagawa

Constantin Franziskus Dominik Hubmann

Haoye Cai

Boxiao Pan

T Wolfram Burgard

German Ros

Bertrand Douillard

De-An Huang

A method for semantic keypoint detection is described. The method includes linking, using a keypoint graph neural network (KGNN), semantic keypoints of an object within a first image of a video stream into a 2D graph structure corresponding to a category of the object. The method also includes embedding descriptors within the semantic keypoints of the 2D graph structure corresponding to the category of the object. The method further includes tracking the object within subsequent images of the video stream using the embedded descriptors within the semantic keypoints of the 2D graph structure corresponding to the category of the object.

End-to-end monocular 2D semantic keypoint detector and tracker learning

A method for vehicle prediction, planning, and control is described. The method includes separately encoding traffic state information at an intersection into corresponding traffic state latent spaces. The method also includes aggregating the corresponding traffic state latent spaces to form a generalized traffic geometry latent space. The method further includes interpreting the generalized traffic geometry latent space to form a traffic flow map including current and future vehicle trajectories. The method also includes decoding the generalized traffic geometry latent space to predict a vehicle behavior according to the traffic flow map based on the current and future vehicle trajectories.

Traffic signal understandings and representation for prediction, planning, and control

Systems, methods, and other embodiments described herein relate to a multi-task model that integrates recurrent models to improve handling of multi-sweep inputs. In one embodiment, a method includes acquiring sensor data from multiple modalities. The method includes separately encoding respective segments of the sensor data according to an associated one of the different modalities to form encoded features using separate encoders of a network. The method includes accumulating, in a detector, sparse features associated with sparse sensor inputs of the multiple modalities to densify the sparse features into dense features. The method includes providing observations according to the encoded features and the sparse features using the network.

Network for multisweep 3D detection

A method for controlling an ego vehicle in an environment includes associating, by a velocity model, one or more objects within the environment with a respective velocity instance label. The method also includes selectively, by a recurrent network of the taillight recognition system, focusing on a selected region of the sequence of images according to a spatial attention model for a vehicle taillight recognition task. The method further includes concatenating the selected region with the respective velocity instance label of each object of the one or more objects within the environment to generate a concatenated region label. The method still further planning a trajectory of the ego vehicle based on inferring, at a classifier of the taillight recognition system, an intent of each object of the one or more objects according to a respective taillight state of each object, as determined based on the concatenated region label.

Vehicle taillight recognition based on a velocity estimation

A method controlling an ego vehicle in an environment includes determining, via a flow model of a parked vehicle recognition system, a flow between a first representation of the environment and a second representation of the environment. The method also includes determining, via a velocity model of the parked vehicle recognition system, a velocity of a vehicle in the environment based on the flow. The method further includes determining, via a parked vehicle classification model of the parked vehicle recognition system, the vehicle is parked based on the velocity of the vehicle and one or more of features associated with the vehicle and/or the environment. The method still further includes planning a trajectory of the ego vehicle based on determining the vehicle is parked.

Parked car classification based on a velocity estimation

A method for an end-to-end boundary lane detection system is described. The method includes gridding a red-green-blue (RGB) image captured by a camera sensor mounted on an ego vehicle into a plurality of image patches. The method also includes generating different image patch embeddings to provide correlations between the plurality of image patches and the RGB image. The method further includes encoding the different image patch embeddings into predetermined categories, grid offsets, and instance identifications. The method also includes generating lane boundary keypoints of the RGB image based on the encoding of the different image patch embeddings.

End-to-end learned lane boundary detection based on a transformer

A method for generating a dense light detection and ranging (LiDAR) representation by a vision system includes receiving, at a sparse depth network, one or more sparse representations of an environment. The method also includes generating a depth estimate of the environment depicted in an image captured by an image capturing sensor. The method further includes generating, via the sparse depth network, one or more sparse depth estimates based on receiving the one or more sparse representations. The method also includes fusing the depth estimate and the one or more sparse depth estimates to generate a dense depth estimate. The method further includes generating the dense LiDAR representation based on the dense depth estimate and controlling an action of the vehicle based on identifying a three-dimensional object in the dense LiDAR representation.

Shared vision system backbone

In one embodiment, a signal light state detection system includes one or more processors, an a non-transitory memory module storing computer-readable instructions. The computer-readable instructions are configured to cause the one or more processors to receive a first image of a vehicle and receiving a second image of the vehicle, wherein the second image is later in time from the first image, and generate a warped image from the first image and the second image, wherein the warped image has individual pixels of one of the first image and the second image that are shifted to locations of corresponding pixels of the other of the first image and the second image. The one or more processors further generate a difference image from the warped image and one of the first image and the second image, and determine, using a classifier module, a probability of a state of vehicle signal lights.

Vehicles, systems and methods for automatically detecting a state of signal lights of a vehicle

System, methods, and other embodiments described herein relate to detection of traffic lights corresponding to a driving lane from views captured by multiple cameras. In one embodiment, a method includes estimating, by a first model using images from multiple cameras, positions and state confidences of traffic lights corresponding to a driving lane of a vehicle. The method also includes aggregating, by a second model, the state confidences and a multi-view stereo composition from geometric representations associated with the positions of the traffic lights. The method also includes assigning, by the second model according to the aggregating, a relevancy score computed for a candidate traffic light of the traffic lights to the driving lane. The method also includes executing a task by the vehicle according to the relevancy score.

Systems and methods for detecting traffic lights corresponding to a driving lane

System, methods, and other embodiments described herein relate to improving the detection of traffic lights associated with a driving lane using a camera instead of map data. In one embodiment, a method includes estimating, from an image using a first model, depth and orientation information of traffic lights relative to a driving lane of a vehicle. The method also includes computing, using a second model, relevancy scores for the traffic lights according to geometric inferences between the depth and the orientation information. The method also includes assigning, using the second model, a primary relevancy score for a light of the traffic lights associated with the driving lane according to the depth and the orientation information. The method also includes executing a control task by the vehicle according to the primary relevancy score and a state confidence, computed by the first model, for the light.

Systems and methods for detecting traffic lights of driving lanes using a camera and multiple models

System, methods, and other embodiments described herein relate to accurately distinguishing a traffic light from other illuminated objects in the traffic scene and detecting states using hierarchical modeling. In one embodiment, a method includes detecting, using a machine learning (ML) model, two-dimensional (2D) coordinates of illuminated objects identified from a monocular image of a traffic scene for control adaptation by a control model. The method also includes assigning, using the ML model, computed probabilities to the illuminated objects for categories within a hierarchical ontology of environmental lights associated with the traffic scene, wherein one of the probabilities indicates existence of a traffic light instead of a brake light in the traffic scene. The method also includes executing a task by the control model for a vehicle according to the 2D coordinates and the computed probabilities.

Systems and methods for detecting traffic lights using hierarchical modeling

Systems, methods, and other embodiments described herein relate to improving trajectory forecasting in a device. In one embodiment, a method includes, in response to receiving sensor data about a surrounding environment of the device, identifying an object from the sensor data that is present in the surrounding environment. The method includes determining category probabilities for the object, the category probabilities indicating semantic classes for classifying the object and probabilities that the object belongs to the semantic classes. The method includes forecasting trajectories for the object based, at least in part, on the category probabilities and the sensor data. The method includes controlling the device according to the trajectories.

Systems and methods for trajectory forecasting according to semantic category uncertainty

A vehicle light classification system captures a sequence of images of a scene that includes a front/rear view of a vehicle with front/rear-side lights, determines semantic keypoints, in the images and associated with the front/rear-side lights, based on inputting the images into a first neural network, obtains multiple difference images that are each a difference between successive images from among the sequence of images, the successive images being aligned based on their respective semantic keypoints, and determines a classification of the front/rear-side lights based at least in part on the difference images by inputting the difference images into a second neural network.

Systems and methods for vehicle light signal classification

Systems and methods described herein relate to jointly training a machine-learning-based monocular optical flow, depth, and scene flow estimator. One embodiment processes a pair of temporally adjacent monocular image frames using a first neural network structure to produce an optical flow estimate and to extract, from at least one image frame in the pair of temporally adjacent monocular image frames, a set of encoded image context features; triangulates the optical flow estimate to generate a depth map; extracts a set of encoded depth context features from the depth map using a depth context encoder; and combines the set of encoded image context features and the set of encoded depth context features to improve performance of a second neural network structure in estimating depth and scene flow.

Systems and methods for jointly training a machine-learning-based monocular optical flow, depth, and scene flow estimator

Growing community of inventors

San Jose, CA, United States of America

Kuan-Hui Lee