Self-supervised learning for video analytics
AssigneeCisco Technology, Inc.
InventorsHugo Latapie, Ozkan Kilic, Adam James Lawrence, Gaowen Liu, Ramana Rao V. R. Kompella
Life of this patent
Estimating the term…
Abstract
In one embodiment, a device represents spatial characteristics of an object depicted in video data over time as one or more timeseries. The device associates different portions of the one or more timeseries with behavioral regimes of the object. The device generates ground truth labels for frames of the video data based on changes in the behavioral regimes of the object associated with those frames. The device trains a self-supervised model to detect an event depicted in the video data using the ground truth labels and their associated frames.
Claims20
A method comprising:
- representing, by a device, spatial characteristics of an object depicted in video data over time as one or more timeseries;
- associating, by the device, different portions of the one or more timeseries with behavioral regimes of the object;
- detecting, by the device, an event in the video data based on a rate of change of the behavioral regimes of the object;
- generating, by the device, ground truth labels for frames of the video data based on the rate of change of the behavioral regimes of the object;
- and
- training, by the device using a machine learning training framework executing on one or more processors, a self-supervised model using the ground truth labels and their associated frames.
An apparatus, comprising:
- a network interface to communicate with a computer network;
- a processor coupled to the network interface;
- and
- a memory configured to store instructions that, when executed by the processor, configure the processor to:
- represent spatial characteristics of an object depicted in video data over time as one or more timeseries;
- associate different portions of the one or more timeseries with behavioral regimes of the object;
- determine a rate of change of the behavioral regimes of the object;
- generate ground truth labels for frames of the video data based on the rate of change of the behavioral regimes of the object;
- and
- train, using a machine learning training framework, a self-supervised model to detect an event depicted in the video data using the ground truth labels and their associated frames.
A tangible, non-transitory, computer-readable medium storing program instructions that cause a device to execute a process comprising:
- representing, by the device, spatial characteristics of an object depicted in video data over time as one or more timeseries;
- associating, by the device, different portions of the one or more timeseries with behavioral regimes of the object;
- determining a rate of change of the behavioral regimes of the object;
- generating, by the device, ground truth labels for frames of the video data based on the rate of change of the behavioral regimes of the object associated with those frames;
- and
- training, by the device using a machine learning training framework executing on one or more processors, a self-supervised model to detect an event depicted in the video data using the ground truth labels and their associated frames.
Dependencies are inferred from the claim text.