Amazon Technologies, Inc.

Kalidas Yeturu

Howard Lee Huddleston, Jr

Sriraghavendra Ramaswamy

Gururaj Narayanan

Nilesh Mahadeo Sah

Anaiappan Govindan

Santhoshkumar Manikandasadanam Kalidasa Varier

An unmanned aerial vehicle (UAV) may perform a surveillance action at a property of an authorized party. The property may be defined by a geo-fence, which may be a virtual perimeter or boundary around a real-world geographic area. The UAV may image the property to generate surveillance images, and the surveillance images may include image data of objects inside the geo-fence and image data of objects outside the geo-fence. While gathering surveillance images, or after the surveillance images have been gathered, the geo-fence information may be used to obscure or remove image data referring to objects outside the geo-fence. Geo-clipped surveillance images may be generated by physically constraining a sensor of the UAV, by performing pre-image capture processing, or post-image capture processing. Geo-clipped surveillance images may be limited to authorized property, so privacy is ensured for private persons and property.

Controlling a route based on priority levels associated with delivery action or surveillance action

A regression model is generated to map observation records of a first dimensionality to a second dimensionality. Using a set of transformed records obtained from the first regression model, a Gaussian mixture model of the distribution of observation records of the second dimensionality is trained. Using a Gaussian distribution obtained from the Gaussian mixture model, a recommended modification of a proposed training set of a classifier is obtained.

Artificial intelligence system for enhancing data sets used for training machine learning-based classifiers

To prevent typos in electronic documents, quality checks are performed to identify portions of the document that have been misspelled. Major sources of false positives are words that have been transliterated from other languages into English. The quality control system described herein identifies transliterated words in electronic documents. The system may be trained to evaluate a set of words, and determine likelihoods that individual words in the set of words are transliterated (i.e., they are not native to the language and/or alphabet of the electronic document). In some embodiments, the model may be further configured to identify a language and/or alphabet that the individual word was transliterated from.

Identification of source languages for terms

Image creation using geo-fence data

The delivery of a package using an unmanned aerial vehicle (UAV) may include performing a surveillance action at a property of an authorized party. The surveillance action may include imaging the property, and may be performed before or after the package is delivered. Surveillance data such as image data gathered in a surveillance action may be modified to remove or obscure adjacent properties not associated with the property of the authorized party. Based on surveillance data, the probability or confidence value of a surveillance event may be determined and a surveillance alert may be generated. Surveillance actions may be scheduled by a central controller to ensure that the surveillance action may be performed without exhausting the resources, such as battery power or time, of the UAV. Surveillance actions may be received as an interrupt and may have an associated priority level.

Unmanned aerial vehicle based surveillance as a service

Digital content may be processed to determine a set of containers in the content. Each container may correspond to a particular text element of the digital content such as a line of text on a page of a digital content file. Container data indicating values of base content properties for each container may be obtained. Derived content properties may be determined from the base content properties and values of the derived content properties may be determined for each container. Multiple iterations of a clustering algorithm may be executed, where each iteration involves grouping the containers into a set of clusters by applying a particular distance function to the values of a particular set of base and/or derived properties for each container. The distance function and set of properties utilized at each iteration may be configurable to obtain clusters that can be associated with particular semantic classifiers.

Multi-level clustering for associating semantic classifiers with content regions

Imagery may be generated for a written work based on textual descriptions in the written work using an automated or partially automated process. The imagery may be used to create a trailer for the written work, for insertion into a digital version of the written work, to visually represent the written work, and/or for other reasons. Descriptive text from the written work may be identified and associated with a catalog of imagery, which include image objects that are discrete images of subjects. The image objects may be compiled with other image objects to create an initial image. The initial image may be edited by a user or other person to add more detail, modify subjects, add an additional subject, remove subjects, change attributes, and/or make other changes to the initial image.

Growing community of inventors

Chennai, India

Kalidas Yeturu