Digital Barriers Services Ltd

University of Strathclyde

Robert Bartholomew Lambert

William Paul Cockshott

Kieron John Messer

Richard John Fryer

Gordon Innes

Andrew Didsbury

An object recognition enabled, for example facial recognition enabled, video surveillance system for capturing video of a scene and allowing recognition of objects within that scene. The system comprises at least one camera apparatus connected via a communication channel to a central server with the camera apparatus arranged for capturing visual representation data of a scene. The visual representation data comprises video of the scene and the camera apparatus comprises a camera for capturing said video and a video encoder for sending corresponding video data via the communication channel to the central server. The camera apparatus is further arranged for generating object recognition data based on said visual representation data, and the video encoder is arranged to send said object recognition data along with the video data via the communication channel.

Video surveillance and object recognition

A video encoder () is disclosed for sending video data to a server () over a channel. A video compressor () can compress video data using an adjustable compression factor. A multiplexer/pack constructor () can receive compressed video from the compressor, together with non-video data, and can queue data for transmission. A bandwidth monitor () is provided for monitoring the performance of the channel based on the behaviour of transmitted data packets, including their round-trip travel time to and from the server (). The compressor () can modify the adjustable compression factor based on the performance of the channel, as determined by the bandwidth monitor, and the amount of non-video data received by the packet constructor. This can permit transmission of non-video data without disrupting the live video feed or introducing any delay.

Video encoder

A method of compressing an array of data entries in the form of digital electronic signals, the method including determining a plurality of possible compressed encodings for each block of the array using a respective different compression technique for each encoding of the block, wherein each possible compressed encoding has a respective data size and a respective reconstruction error; and generating a list of the plurality of compressed encodings and selecting that one of the plurality of compressed encodings from the list which, when uncompressed and added to the reference data array, provides the largest relative improvement in reconstruction error per unit increase in data size. A compressed encoding of the original data array may then be updated by adding the selected compressed encoding, and a reference data array may be updated by adding the selected encoding in uncompressed form. These steps may be different combinations of macro block regions repeated recursively until a maximum data size is achieved for the compressed encoding of the original data array.

Data compression

A recognition system of the self-organizing artificial neural network type is arranged to classify input data according to stored categories which have been determined by a training process. In the training process the initial category representations are selectively iteratively updated in response to a series of training patterns and in accordance with a competitive learning routine. this routine uses measures of category utilization based on the proportion of all inputs received over a representative period, particularly long term utilisation and short term utilization, to ensure that all available categories will be used and that the system is stable. The training rate which determines the amount of modification to a category representation at an up-date is local to each category and is based upon the maturity of the category and on the similarity measure between the internal representative pattern and the training input so that the training duration can be minimized. A user-operated selectively-operable suggestion learning input is provided to each category to modify the training process or to enable secondary training to proceed during classification of input data using that input data as the training patterns. The categories are represented by multiple reference patterns with respective importance values from which the degree of compatibility between an input and a category is computed taking into account the importance values.

Recognition system

A data compression system using vector quantization utilises a codebook or index tables constructed by finding a small set of points in an n-dimensional space which are representative of a much larger population. The space is partitioned such that the integrals of the probability density functions over all the subspaces are approximately equal and thereafter the centre of gravity of each subspace is taken to represent the subspace population. The partitioning is effected sequentially to map n-dimensional space into one dimensional code space.

Growing community of inventors

London, United Kingdom

Robert Bartholomew Lambert