Abstract: Methods, systems, apparatuses, and computer program products are provided that are configured to processor sensor data using a single component specialized sensor data processing neural network. The single component data processing system is configured to receive input data from a sensor, analyze the input data using a neural network embodied by a single trained sensor data processing component, wherein the single trained sensor data processing component is trained to product output data that approximates a plurality of task-specific transformations performed in a pipeline manner, and produce the output data following transformation of the input data. Additionally, methods, systems, apparatuses, and computer program products are provided for training a single trained sensor data processing component to approximate a plurality of task-specific transformations performed in a pipeline manner.

Abstract: A method, apparatus, and computer program product are provided to determine sensor orientation. In the context of a method, the orientation of a sensor is predicted based on a predetermined dataset. The method also includes receiving information regarding a location of the sensor and information from which a gravity direction of the sensor is derivable. The method further includes determining a relative orientation of the sensor in relation to a frame of reference and the gravity direction of the sensor. The method still further includes comparing the orientation of the sensor that was predicted and the relative orientation of the sensor to determine a prediction accuracy. The method finally includes updating the predetermined dataset based on the prediction accuracy of the sensor orientation.

Abstract: Methods described herein relate to reducing the computational intensity of vision-based localization. Methods may include: receiving sensor data from a vehicle traveling along a road; identifying one or more features of the environment from the sensor data; classifying the one or more identified features into one or more of a plurality of semantic classifications for the features; identifying map image data based on an identified location of the vehicle; identifying one or more features in the map image data; comparing one or more identified features of a first semantic classification with one or more features of the map image data of the first semantic classification; and registering a localized location of the vehicle within the environment based, at least in part, on the one or more identified features of the first semantic classification corresponding to the one or more features of the map image data of the first semantic classification.

Abstract: A method, apparatus and computer program product are provided for warping a perspective image into the ground plane using a homography transformation to estimate a bird's eye view in real time. Methods may include: receiving first sensor data from a first vehicle traveling along a road segment in an environment, where the first sensor data includes perspective image data of the environment, and where the first sensor data includes a location and a heading; retrieving a satellite image associated with the location and heading; applying a deep neural network to regress a bird's eye view image from the perspective image data; applying a Generative Adversarial Network (GAN) to the regressed bird's eye view image using the satellite image as a target of the GAN to obtain a stabilized bird's eye view image; and deriving values of a homography matrix between the sensor data and the established bird's eye view image.

Abstract: A method, apparatus and computer program product are provided for establishing the reliability of crowd sourced data based on the source of the data and the region in which the data is gathered. Methods may include: receiving map data for a network of roads in a geographic area; receiving location accuracy data from different regions within the geographic area, where the location accuracy data for each region may be indicative of an accuracy with which location can be established in the respective region; receiving a plurality of probe data points; determining, for each probe data point, the location accuracy data associated with the location information associated with the respective probe apparatus; determining, for each probe apparatus, a reliability of the one or more sensors; and updating a map database based on at least one probe data point.

Abstract: An approach is provided for providing predictive classification of sensor error. The approach involves, for example, receiving sensor data from at least one sensor, the sensor data collected at a geographic location. The approach also involves extracting a set of input features from the sensor data, map data representing the geographic location, or combination thereof. The approach further involves processing the set of input features using a machine learning model to calculate a predicted sensor error of a target sensor operating at the geographic location. The machine learning model, for instance, has been trained on ground truth sensor error data to use the set of input features to calculate the predicted sensor error.

Abstract: Various methods are provided for facilitating map update to an environment map using gradient thresholding. One example method may include detecting an observed feature associated with a first feature decay and generating an interpolated feature that approximates the observed feature associated with a second decay. The method also includes determining a gradient difference between the interpolated feature and a stored map feature. The stored map feature represents an encoding of the observed feature associated with a third decay associated with an environment map. The method also includes determining a relationship between the gradient difference and a feature gradient update threshold, and, based upon the relationship, updating the environment map by at least replacing the map feature representation associated with the environment map with the approximated feature representation.

Abstract: Described herein are methods of generating learning data to facilitate de-biasing the labeled location of an object of interest within an image. Methods may include: receiving sensor data, where the sensor data is a first image; determining reference corner locations of an object in the first image using image processing; generating observed corner locations of the object in the first image from the determined reference corner locations; generating a bias transformation based, at least in part, on a difference between the reference corner locations and the observed corner locations of the object in the first image; receiving sensor data from another image sensor of a second image; receiving observed corner locations of an object in the second image from a user; and applying the bias transformation to the observed corner locations of the object in the second image to generate de-biased corners for the object in the second image.

Abstract: Described herein are methods of detecting and labeling features within an image of an environment. Methods may include: receiving sensor data from an image sensor, where the sensor data is representative of a first image including an aerial view of a geographic region; detecting, using a perception module, at least one vehicle within the image of the geographic region; identifying an area around the at least one vehicle as a road segment in response to detecting the at least one vehicle; based on the identification of the area around the vehicle as a road segment, identifying features within the area as road features based on a context of the area; generating a map update for the road features of the road segment; and causing a map database to be updated with the road features of the road segment.

Abstract: An approach is provided for fully-automated learning to match heterogeneous feature spaces for mapping. The approach involves determining a first feature space comprising first features and a second feature space comprising second features, and classified by a feature detector into a first attribution category and a second attribution category, respectively. The approach further involves calculating a first similarity score for the first feature space based on a first distance metric applied to the first features, and a second similarity score for the second feature space based on a second distance metric applied to the second features. The approach also involves determining a transformation space comprising a first weight to be applied to the first similarity score and a second weight to be applied to the second similarity score based on matching the first attribution category and the second attribution category.

Abstract: Methods, systems, apparatuses, and computer program products are provided that are configured to perform localization of position data, specifically using a trained localization neural network. In the context of an apparatus, the apparatus is caused to receive observed feature representation data. The apparatus is further configured to transform the observed feature representation data into standardized feature representation data utilizing a trained localization neural network. The apparatus is further configured to compare the standardized feature representation data and the map feature representation data and identify local position data based on the comparison.

Abstract: Methods described herein relate to iteratively refining parameters of a localization framework for different data sources. Methods may include: receiving a map data set from a data source; applying an offset to the map data to generate an offset map data set; applying a localization framework between the map data set and the offset map data set to generate a localized data set; recovering offset data from the localized data set; iteratively tuning parameters of the localization framework in response to the offset data from the localized data set failing to correspond with the applied offset until offset data recovered from the localized data set corresponds to the applied offset; receiving sensor data from at least one sensor of a vehicle; applying the localization framework to the sensor data to locate the vehicle within a mapped region.

Abstract: A method is provided for dynamically throttling processing and memory consumption of sensors based on context. Methods may include: receiving first sensor data from a first sensor, where the first sensor data includes data associated with an environment of the sensor; applying an auto-encoder framework to the first sensor data to establish a data difference score between frames of the first sensor data, where a relatively high data difference score corresponds to substantial differences between frames of the first sensor data, and wherein a relatively low data difference score corresponds to insubstantial differences between frames of the first sensor data; reducing a frame rate of data capture of the first sensor in response to the data difference score between frames of the first sensor data being relatively low; capturing first sensor data from the first sensor at the reduced frame rate; and providing for storage of the first sensor data.

Abstract: A method is provided for generating training data to facilitate automatically locating an object of interest within an image. Methods may include: receiving sensor data including a plurality of images from at least one image sensor; receiving an identification, from a user, of an object visible within an image of the plurality of images, where at least a portion of the object is visible in one or more of the plurality of images; determining a predicted location of the object in the one or more of the remaining images of the plurality of images; identifying the object in the one or more of the remaining images of the plurality of images; and storing the plurality of images including an indication of the object at the object location within the one or more of the plurality of images.

Abstract: An approach is provided for storing and retrieving map data using contextual information priors. The approach involves, for example, processing contextual information to determine a restricted range of location information relevant to at least one query. The approach also involves processing sensor data received from at least one sensor, the sensor data collected at at least one query location, to determine semantic information. The approach further involves filtering the map data based, at least in part, on the restricted range of location information relevant to the at least one query, the semantic information, or a combination thereof.

Abstract: An approach is provided for localizing a vehicle pose on a map. The approach involves, receiving an input specifying the vehicle pose with respect to a road lane of the map. The approach also involves searching over a set of candidate lateral offsets to select a lateral offset that minimizes a lateral error between the vehicle position with the lateral offset applied and a lateral location of the road lane, wherein the lateral location and the travel direction of the lane are determined from the map. The approach further involves searching over a set of candidate vehicle headings at the selected lateral offset to select a vehicle heading that minimizes a heading error. The approach further involves determining a local optimum of the vehicle pose based on the selected lateral offset and vehicle heading, wherein the vehicle pose is localized to the map based on the local optimum.

Abstract: A method, apparatus and computer program product are provided for predicting the likelihood that road infrastructure has changed. Methods may include: receiving sensor data from a sensor network of a first environment; identifying first map data of an environment, where the first map data is identified based on visual similarity to the first environment, where the first map data includes a history of map updates; identifying second map data of an environment, where the second map data is identified based on visual similarity to the first environment; analyzing the sensor data against the first map data and the second map data to establish correspondence between the sensor data and the first map data and between the sensor data and the second map data; and identifying the first environment as a location of at-risk infrastructure based on the sensor data corresponding to the first map data.

Abstract: A method, apparatus and computer program product are provided for constructing a high definition map from crowd sourced data using semantic attributes to bootstrap map construction. Methods may include: receiving first sensor data from a first vehicle having traversed a first path along a first lane of a first road segment; identifying features from the first sensor data of the first road segment; receiving second sensor data from a second vehicle having traversed a second path along a second lane of the first road segment; identifying features from the second sensor data of the first road segment; aligning the identified features from the second sensor data with the identified features from the first sensor data of the first road segment; and combining the identified features from the first sensor data and the second sensor data based, at least in part, on the confidence of the respective sensor data.

Abstract: An approach is provided for generating a reverse sequence or real-time streamed images for triangulation. The approach includes receiving a real-time stream of images captured by a sensor of a vehicle during a drive; extracting a sequence of two or more images from the real-time stream; reversing the sequence of the two or more images; and providing the reversed sequence of the two or more images for feature triangulation.

Abstract: An apparatus, method and computer program product are provided for predicting feature space decay using variational auto-encoder networks. Methods may include: receiving a first image of a road segment including a feature disposed along the road segment; applying a loss function to the feature of the first image; generating a revised image, where the revised image includes a weathered iteration of the feature; generating a predicted image using interpolation between the image and the revised image of a partially weathered iteration of the feature; receiving a user image, where the user image is received from a vehicle traveling along the road segment; correlating a feature in the user image to the partially weathered iteration of the feature in the predicted image; and establishing that the feature in the user image is the feature disposed along the road segment.