Systems and methods for detecting anomalies from data
The present disclosure concerns methods and/or systems for processing, detecting and/or notifying for the presence of anomalies or infrequent events from data. Some of the disclose methods and/or systems may be used on large-scale data sets. Certain applications are directed to analyzing sensor surveillance records to identify aberrant behavior. The sensor data may be from a number of sensor types including video and/or audio. Certain applications are directed to methods and/or systems that use compressive sensing. Certain applications may be performed in substantially real time.
1. A method for processing, detecting and/or notifying for the presence of at least one infrequent event from at least one large scale data set comprising:
receiving time series data;
representing either the time series data, or one or more features of the time series data, as sets of vectors, matrices and/or tensors;
performing compressive sensing on at least one vector, matrix and/or tensor set;
decomposing the at least one compressive sensed vector, matrix and/or tensor set to extract a residual subspace; and
identifying, using a computing device, potential infrequent events by analysing compressive sensed data projected into the residual subspace.
2. A method according to claim 1 , wherein the one or more features of the time series data are one or more variables associated with a subset of the time series data.
3. A method according to claim 1 , wherein the one or more features of the time series data are the difference in any variable between any two subsets of the data of the time series data.
4. A method according to claim 3 , wherein the time series data is a stream of at least one of video and audio surveillance data, each subset is a frame and the feature is a motion or frequency information between each successive pair of frames.
5. A method according to claim 3 , wherein the time series data comprises a stream of video data, each subset comprises a frame and the feature comprises motion information between each successive pair of frames, the motion information being determined by dividing each frame of video by a grid into cells, and then counting a number of optic flow vectors in each cell to provide optical flow information.
6. A method according to claim 5 , wherein the optical flow information is represented using bag-of-visual-words.
7. A method according to claim 6 , wherein from the bag-of-visual-words a feature-frame matrix is constructed by amalgamating the feature vectors for all frames in the sequence.
8. A method according to claim 7 , wherein the feature-frame matrix is then structurally decomposed into the observed into principal and residual components.
9. A method according to claim 8 , further comprising detecting abnormal events in the residual subspace using a Q-statistic based test statistic.
10. A method according to claim 1 , wherein a pre-cursor step is employed to transform the data to the Compressed Domain, to reduce the data to a manageable size.
11. A method according to claim 10 , wherein the pre-cursor step of transforming the data to the compressed domain is deployed for reducing a feature dimension in the compressed domain and employs a sensing matrix having entries with values of either 0 with probability 2/3 or +/−1 with probability 1/6.
12. A method according to claim 11 , wherein a random gossip algorithm is applied.
13. A method according to claim 10 , wherein the pre-cursor step of transforming the data to the compressed domain is deployed for reducing a time instances in the compressed domain and employs frame sub-sampling.
14. A method according to claim 10 , wherein the features are represented as sets of vectors before being transformed to the Compressed Domain.
15. A method according to claim 1 , wherein said identifying potential infrequent events involves thresholding the data projected into the residual sub-space.
16. A system comprising:
at least one sensor to receive data comprising a time series of data subsets for analysis;
a computer memory for storing the data;
a computer processor for:
representing either the time series data, or one or more features of the time series data, as sets of vectors, matrices and/or tensors;
performing compressive sensing on the at least one vector, matrix and/or tensor set;
decomposing the selected sets of compressive sensed vectors, matrices and/or tensors to extract a residual subspace; and
identifying potential infrequent events by analysing compressive sensed data projected into the residual subspace.
17. The system of claim 16 , wherein the computer processor is co-located with the at least one sensor.
18. The system of claim 16 , wherein the computer processor is remotely located from the at least one sensor.
19. The system of claim 16 , wherein a portion of the computer processor is remotely located from the least one sensor and performs the step of identifying potential infrequent events by analysing compressive sensed data projected into the residual sub-space.
20. The system of claim 16 , wherein the functionality of the computer processor is divided between a sub-processor co-located with the at least one sensor and a sub-processor remotely located from the at least one sensor.