A brain for digital eyes

Abstract

xCalibre is an advanced, real-time, image recognition system. It offers businesses, governments, militaries, and private clients smart detection and recognition functions with CCTV, recorded videos, and images. xCalibre makes it possible to collect and analyze information in real-time, identify humans and objects, and send alerts in defined cases. The system is AI-based and can make predictions and recommendations for optimization based on real-time statistics and historical data. xCalibre identifies humans, with or without facial and body covers, and concealed objects as well. The system includes an advanced cybersecurity protocol to keep all data secure and confidential.

xCalibre is an advanced image-analysis technology, with intelligent detection and recognition functions, that allows a wide array of users to enhance the capabilities of real-time video and CCTV systems. xCalibre makes it possible to analyze real-time cameras, recorded video, and images and identify humans and objects for further processing. The system optimizes data and, using its neural network-based analysis, builds a database, keeps history, and enables crucial security alerts – in real time of course.

The system’s algorithms are point detector-based. They can identify points in an image with 2D changes. A geometric-feature evaluator overlays at least one mesh on an image and analyzes features on at least one mesh. An internal calibrator transforms data from the point detector and the geometric-feature evaluator into a 3D point figure of the image. A depth evaluator determines final shape. A 3D object model of the image (face, body, or object) is constructed. The system can construct and learn features on a partial view, e.g. where a face is to some extent covered.

xCalibre includes an AI analysis that works in phases. In the first phase, a high resolution, pixilation-based image mapping is performed using a neural network and associated neural-network analysis. A second phase uses a proprietary convolutional neural network (CNN) named First Encounter (FECNN). First

Encounter performs low-resolution, pixilation-based mapping of an image (face body, or object). In a third phase, portrait-wide mapping is performed using the neural network. In phase four, sideways mapping using an expert system can also be run. In phase five, a biometric facial-mapping is performed using the expert system. Phase six entails a human body style study based on AI vector mapping. Finally, phase seven, based on AI-vector mapping, identifies clothing, face covering, and accessories. A person-of-interest (or object) can be identified. For humans, a complete identification can be done with or without clothing or face cover.

xCalibre’s intelligent imaging offers a new generation of surveillance. GBT’s patented First Encounter image-processing algorithm is a remarkable breakthrough. Based on innovative media processing techniques, our First Encounter algorithm performs rapid analyses of huge quantities of data.

xCalibre’s advanced, AI-surveillance analysis will be an efficient solution for commercial, government, military, and all security markets.

xCalibre can identify and analyze vehicles and humans

xCalibre can tell the difference between objects and humans

xCalibre technology works with global video and imaging via a private, secure network or the cloud. Its sophisticated machine learning enables surveillance cameras, including natural language (NLP) commands. It works with real-time CCTV system or live stream video. xCalibre is highly scalable. That permits quick and easy adaptation for size and capacity. It supports commercial and individual setups, live-stream data analysis, and storage of vital information such as location, time, and logs.

One of the key advantages of our xCalibre system is its ability to differentiate and

categorize events, humans, and objects. xCalibre can tell the difference between a speeding vehicle and a car accident, alerting the appropriate authorities in real time. The system records vital details for future usage. For example, if a person-of-interest is detected, the system relays timely information to authorities and records the event for future reference.

xCalibre includes advanced AI features such as natural language processing (NLP) to search for details of specific persons or objects. A verbal search can be done globally on live or archived surveillance video in the following fashion:

“Show me footage of a woman wearing a red coat between these dates at Dallas-Fort Worth Airport.” Or: “Alert for a tall man with a black suitcase who crosses the corner of Main and Syracuse Streets in the next hour.” The system will go to work instantly.

xCalibre recognizes human features

xCalibre measures and recognizes height, width, depth, and unique features of any person or object. It also supports advanced facial recognition that deals effectively with facial coverings. The tool maintains all information in a central data center from which it can send notifications for defined individuals and objects. Our technology can identify a person in full clothing or with facial covering. Objects of interest (e.g., concealed weapons) can be identified and notifications can go out instantly.

xCalibre is equipped with facial emotional analysis to identify and predict scenarios. This type of analysis can warn of a possible attack or intentions to commit crime. The technology can be set up to perform as a personal or global security watchdog to assist live agents with suspicious activities. The capability

to identify hidden objects on humans (e.g., firearms,) makes it an intelligent tool for use in public places like airports, sport venues, concert halls, and demonstrations. The system can perform as a crowd “visual searcher” and alert in case of concealed objects. Working with video streams, it can analyze what’s happening in real time. A hidden live “visual search” can be extraordinarily useful in scanning students and other people entering our schools and places of worship. Individuals who sweat or whose pupils indicate stress can be readily detected – and of course alerts will be transmitted instantly to responsible parties.

xCalibre’s AI technology learns facial and body features such as skull size, distance between eyes, and bone structure. The system includes hardware, a method of processing, and software that operates on multi-core CPUs to achieve real-time effectiveness. It also includes mobile app materials.

Our detection and identification algorithm finds specific, predefined points in the image and computing descriptors for features around them. These detectors identify points in the image for which the signal changes two-dimensionally – that is, at corners, junctions, and vertexes of faces, body parts, and objects. Computer vision algorithms are applied directly to 3D data to develop detectors for locally interesting points. xCalibre applies vision-based, interest-point detectors on depth images to construct 3D object models, from partial views in an unsupervised fashion. Another proprietary algorithm considers spectral geometric features on triangular meshes and using stereotypical vision, recognizes pointing gestures. Using standard computer vision techniques, imaging produces colored areas to identify the location of the face and body features. The algorithm includes depth evaluator to determine final shape.

xCalibre’s many important features can be pre-set using rule-run set decks in TCL format, which can be manually set up using command line orders.

The main features our system performs:

1. Facial detection, recognition, and categorization. Full human face recognition and matching in the database. xCalibre can detect multiple faces at a time –

say, a few people in a party or on a street or at an airport. The system identifies facial attributes and emotional states, alerting responsible parties in real time if warranted.

2.

Human’s features detection, recognition and analysis. Full-body feature detection, recognition, and categorization. Identification of unique, out of the ordinary features. Complete analysis and storage for future use.

3.

Object detection and identification. Identification of objects according to categories. Recognizing objects-of-interest and alerting in real time.

4.

Detection, analysis and classification of human emotions. Classification of emotional state and body language of humans. Alerts in real time for specific requirements.

5.

Animal detection, recognition, and classification.

6.

Detection of signs, logos, scenery, and landmark detection. Artificial or natural scenery and landmarks are detected and classified.

7.

Explicit content detection and identification. Car accidents, bank robberies, weapon discharges, emergencies, and the like are detected and alerts sent to relevant authorities.

8.

AI-based natural language processing (NLP). Searches video for humans and objects according to defined criteria.

9.

Identifying humans with/without facial/body coverage/clothing. Identifying concealed objects.

10.

Real-time operation. Using computer/web-based control panel synchronized with mobile app.

xCalibre performs facial recognition

xCalibre’s emotion recognition

xCalibre recognizes various objects

Image recognition is well-known these days. It’s on the news, in the movies, and used in a wide variety of spheres. xCalibre’s image-recognition is powered by deep learning, specifically Convolutional Neural Networks (CNN). That’s a neural network architecture that emulates the human visual cortex in order to analyze images.

Another component of xCalibre is an RNN (Recurrent Neural Network) module that drives speech recognition and NLP processing. Our proprietary CNN is based on a unique image-recognition approach. It works in conjunction with our patented core-image-analysis algorithm called First Encounter. Deep-learning algorithms enable real-time detection and identification of humans and objects. This is particularly important for real-time responses with systems like autonomous machines, drones, and medical devices.

FECNN – First Encounter CNN algorithm is the core

FECNN is the core of xCalibre. It’s a dynamic CNN that we painstakingly developed and are proud to introduce. FECNN has cognitive capabilities and is self-adapting to the system’s usage model. For example, if xCalibre is surveilling a shopping mall, its FECNN will adapt functionality according to the mall’s specific topography and structures. For an airport, FECNN adapts to the site’s counter, walkways, luggage carousels, and checkpoints. The algorithm is self-trained and becomes more knowledgeable over time.

The FECNN sees images as 2/3D models, observing movements in picosecond intervals. FECNN scans and analyzes data on-the-fly to identify and extract important features that will then be used to detect and classify. Operating in background, FECNN is used for specific footage search, analysis, human/object detection. It’s constantly referencing similar graphical information to provide accurate conclusions.

FECNN Architecture

In a typical CNN system, neurons in one layer communicate with neurons in the next layer to maintain internal connectivity. With huge quantities of data to analyze, this can become inefficient and slow. If an object has millions of pixels, the number of parameters using a traditional neural network may grow exponentially, leading to data explosion. FECNN architecture is structured with a similar functionality of the visual cortex enabling vast data processing – and fast! Its design includes a heuristic-based, successive approximation approach. The system’s adaptation and cognitive capabilities make it operate much like the human brain.

A typical CNN has a 3D structure. However, the neurons in our FECNN split into a 4D structure. Every set of neurons analyzes a specific aspect of an image. The FECNN uses predictive methods to complete missing parts with sets of computational geometry-based algorithms. These methods classify an image’s features according to identified key pointers and probabilistic approximation.

The end result is a system that quickly learns and adapts to an image’s characteristics. By keeping images in history, the system compares them in real time and reaches instantaneous conclusions. Once an image is encountered, analyzed, and classified, similar ones will be quickly identified. Using its successive completion feature, even if the image is not identical to previous ones,

the system focuses on the different features only. This enables rapid identification. The FECNN is working incrementally, i.e. processing differences and variations. It’s the first of its kind to actually learn images over time, exactly as we humans do.

FECNN processes data incrementally
Reliable, secure, and FAST!

FECNN is a high-performance algorithm that executes each operation several times for error-correction and verification. Redundancy checks ensure accurate detection and identification. Using GBT’s patented database management technology, an image data are classified into objects, such as human, structure, animal, and more. The algorithm records all data and characteristics. FECNN then identifies sub-features within each object. For example, if the object is a car, then contours, wheels, colors, and body trim. In parallel, FECNN analyzes

background features of interest such as scenery and location.

FECNN includes a self-cognitive feature. After analyzing an object, it’s learned and recorded along with its items-of-interest, then stored for future use. If a related object is detected down the road, FECNN will immediately use the relevant data without searching and analyzing. The cognitive part of the FECNN is self-training which provides knowledge growth from every analysis. If a similar condition occurs, FECNN is trained to act immediately. It’s not hard to see how crucial this can be in sensitive places such as our airports.

xCalibre includes an expert system that provides ratios and relation analysis between unique key pointers in faces and bodies. The same holds for inanimate objects. For example, pictures taken without mask/clothing are mapped by the expert system to identify vectorials based on relations and ratios of key features. Unique features are mapped as 2D and 3D databases for later use as key identifiers to recognize people with facial or body coverings. The expert system computes physical relations and ratios of unique facial and body features such as distances, e.g. the distance between nose and mouth on the right side of the face relative to the distance on the left. Other computations include depth level, e.g. the depth between the left eye relative to that of the right.

Each object is scanned to identify key pointers and distances ratios. Expert system input is the FECNN data and the output is a vectorial map of unique features. The process can be done for all facial, body, and object views. The facial and body views are front, back, and sides.

The FECNN operates in a parallel processing method using supervised successive approximation and deep learning-based prediction methods. The algorithm can handle vast amounts of data in real time. xCalibre operates very much as we humans do. That’s why we call it “the brain behind digital eyes.”

FECNN classifies human’s features

GBT is designing a proprietary CLEVER IPU (Intelligence Processing Unit) microchip to drive xCalibre. The company started to design a smart IPU chip to mimic the human brain. The CLEVER chip will be the core brain with a wide variety of tasks in surveillance systems.

The microchip will process immense amounts of data and accelerate the progress of xCalibre within smart applications for IoT, security systems, autonomous cars, communications, medicine, and more.

GBT’s AI relies on neural networks that mimic the network of neurons in human brains. First Encounter CNN, a centerpiece of GBT’s AI system, is based on complex mathematical algorithms so that it can learn by processing vast amounts of data.

In order to drive these sophisticated features, a new kind of computing power is needed. Today, most companies analyze data using graphics processing units. GPUs were originally designed to handle high-resolution graphics for games, but due to their formidable capabilities, they are also used to run mathematical applications, including neural networks.

GBT’s CLEVER is a powerful IPU chip for higher processing power and advanced functionalities, especially running neural networks and related algorithms. Typically, an AI system operates under many integrated circuits working together within an electronic circuit and along with a computer program. Each chip performs a specific task. While processing information, a large amount of data is moving between the chips – and that slows the system. GBT’s CLEVER chip is designed to include all important components onboard, including memory, graph compiler, and other proprietary units. This aids internal processing and saves processing time.

Building all this inside one chip will significantly increase processing speed and reduce power consumption. The CLEVER chip will not be small compared to peers, but its superior performance justifies its size. It will boost xCalibre with higher computing power for neural networks circuitry. The CLEVER IPU will be accompanied by computing software to enable knowledge graphs for representation and reasoning. The combination of knowledge-based algorithms and the CLEVER IPU’s superior power enables the ultimate intelligent application – xCalibre.

xCalibre will be driven by the CLEVER chip

xCalibre algorithms are computationally intensive – indeed highly so. Its proprietary point transformation is key. Our internal calibration algorithm transforms measurements of distances into 3D point figures. It identifies points of interest based on geodesic distance between vertices in a mesh. This provides a stable estimate of a person’s shape and pose that can be used to digitize his/her image prior to construction. Based on these algorithms, xCalibre further categorizes a person’s features and removes facial and body covers, reconstructing the person’s face and body for the most accurate identification.

xCalibre’s deep learning is a set of goal-oriented algorithms that constantly learn about objects in the most efficient way. Deep learning is based on vectorization of face, body, and objects. These algorithms evolve over time, getting more knowledgeable and aware. xCalibre can be installed for a specific purpose and put through self-training time to become an “expert in its field.” For example, installing xCalibre in an airport environment with initial preset rules, and letting it operate for a few weeks, will make it a master of the location’s specifics. Under

FECNN supervision, it will identify people of interest, concealed objects, and more.

xCalibre’s 3D points of interest transformation algorithm

xCalibre’s deep learning matches immediate actions with required scenarios. Its detection and identification algorithms finds specific, predefined points in the image and computing descriptors for features around them. It constantly studies all human/objects within the given media and operates as an active viewer. These methods apply to facial and body part analysis. xCalibre’s computer vision algorithms are applied directly to 3D data to develop models for objects-of-interest. xCalibre’s deep learning algorithm will learn from situations for future use. In a short time, it will perform faster, if similar conditions recur.

The system reacts in real-time, observing, detecting, analyzing, and performing actions. The FECNN algorithm, the core of xCalibre’s technology, is designed as a multi-layered, object-oriented program to get information and perform rapid detection and identification. Based on self-training data, the system can make high-probability predictions and investigate, on its own initiative, within areas it

finds of interest. This enables an intelligent surveillance system for airports, crowded areas, conferences, and demonstrations.

Vectorization is achieved via a 3D points of interest transformation algorithm

xCalibre’s deep learning algorithms learn from real life situations

xCalibre’s algorithms use a proprietary successive-approximation method that checks and validates data against trusted sources. When the algorithm identifies an incorrect conclusion, it adjusts – automatically of course. FECNN algorithm supervises the data, reaches the best conclusion, and takes action.

xCalibre’s AI methods include supervised and unsupervised techniques. Supervised methods are based on training runsets to reproduce results – for example, recognition of specific cars in a live stream of a highway. Another example is identifying a person-of-interest in a dense crowd in a busy airport – and if a match is found, sending an alert. The self-training process is governed by the FECNN algorithm which enables high-speed processing. After training, the system can detect and identify images and send alerts in real time. This supervised process is fast and accurate.

Image source: http://bigdata-madesimple.com

xCalibre also includes an unsupervised system. Here xCalibre estimates hidden

structures based on previous experience and real-life information. An example is identifying a concealed weapon based on out-of-the-ordinary clothing shape. The system identifies an out-of-the-ordinary shape and completes the operation using its own knowledge. If it concludes it may be an object-of-interest, it’ll send an alert.

xCalibre’s unsupervised learning algorithms work on their own to constantly

discover new information. After it detects and learns, it enters a record for future use. xCalibre’s unsupervised part performs complex processing tasks with the

FECNN algorithm. Although unsupervised learning techniques are considered more unpredictable compared with other natural learning methods, we equip xCalibre’s mechanisms with a self-check, supervised algorithm to ensure

accurate and reliable results.

Image source: http://bigdata-madesimple.com

xCalibre includes successive approximation technology to perform multi-scenario management analysis. Quickly identifying objects in a live-streaming video is a formidable challenge. To achieve a significant computational gain, we developed a structured interaction system based on an error-bound algorithm.

xCalibre is designed to work in a dynamic environment. Based on vast numbers of inputs, it needs to quickly detect, identify, analyze, and make decisions. The overall activity must be completed instantly. The algorithm models live events, forming objects’ discrete distribution and mapping them into a multidimensional associative array. While it operates independently, constantly performing sub-analyses per scenario, it receives local feedback for each completed analysis and builds global feedback data.

Once it has sufficient information, the algorithm reaches a conclusion and makes a decision. The algorithm is formulated using error-bound infrastructure to sort out non-relevant data and reach real-time conclusions. The data is fed into a

verification algorithm as another error-check layer. The algorithm’s formulation allows it to handle a great amount of data analysis. The supervised, successive approximation approach provides a near optimal solution which in conjunction with a verification mechanism, enables an optimal conclusion as a basis for decision. The algorithm is scalable and adaptable to a dynamic, multiple scenario environment. It’s crucial to the reliability and accuracy of the system.

xCalibre’s successive approximation provides advanced image mapping

Modern computing has introduced advanced security systems that rely greatly on immense data analysis and operations. However, it’s still a major challenge for reliable surveillance systems. xCalibre introduces a new technology to enable real-time analysis from a large number of cameras, sensors, and other IoT devices. These digital eyes collect data 24/7. xCalibre’s the most advanced solution to handling large volumes of data in a rapid manner, as our world is

demanding more and more. xCalibre is an intelligent, adaptable, self-cognitive system that works with IoT/mobile devices to provide automatic security surveillance for complex locations like airports, conventions, sports venues, performance halls, military situations – and anywhere threats lurk.

xCalibre detects, identifies, and analyzes humans and objects with or without face and body covering. It learns and differentiates between typical and out-of-the-ordinary activities of humans, objects, behavior, or events. All this is done via xCalibre’s proprietary image-analysis processing algorithm – First Encounter Convolutional Neural Network. Our FECNN algorithm is based on innovative media processing technique enabling it to perform rapid analysis of huge amounts of data.

xCalibre is the digital brain behind the digital eyes. It will revolutionize the vital field of surveillance in our dangerous world. In a world filled with danger, conclusions and decisions must be both fast and reliable. Our xCalibre system is up to the task!