Assistive technology

Abstract

There is hereon described assistive and computational technology, including, but not limited to, applications for visually impaired and blind persons, as well as other domains requiring efficient data processing on low-memory and low-compute devices.

Description

[0001] ASSISTIVE TECHNOLOGY

[0002] Field of the Invention

[0003] This invention relates generally to assistive and computational technology, including, but not limited to, applications for visually impaired and blind persons, as well as other domains requiring efficient data processing on low-memory and low-compute devices.

[0004] Background of the invention

[0005] Various assistive technologies exist to help users engage with, and navigate, their environment safely, enabling independent living and participation in activities that might otherwise be challenging. These technologies include, but are not limited to, assistive devices for visually impaired users, scene recognition systems, text recognition and translation systems, fall detection systems and personal assistance systems.

[0006] Scene recognition systems have become increasingly important in a variety of applications, particularly for providing contextual information and enhancing user experience in augmented reality (AR) and assistive technologies. Traditional scene recognition systems often struggle with accuracy and efficiency, especially when deployed on devices with limited computational resources.

[0007] Text recognition and translation systems are increasingly important in a variety of applications, particularly for visually impaired users. Traditional OCR systems often struggle with accuracy in challenging conditions such as low light, distorted text, and diverse fonts. Furthermore, the integration of translation and contextual understanding capabilities is limited in existing solutions.

[0008] Fall detection systems are crucial for ensuring the safety and well-being of individuals, particularly the elderly and those with mobility issues. Traditional fall detection systems often suffer from high false positive rates and require significant computational resources, limiting their effectiveness in low memory and low compute wearable devices.

[0009] Personal assistance systems are crucial for providing timely help and enhancing the safety and convenience of users, particularly those with disabilities or in emergency situations. Traditional systems often require significant computational resources, limiting their effectiveness in low memory and low compute wearable devices.

[0010] Aspects of the present invention seek to address at least one or more of these issues by innovative processing techniques.

[0011] Statements of Invention

[0012] In accordance with a first aspect of the invention, there is provided a method for optimizing input data (including, but not limited to, media, sensor and textual data) and determining relevant features thereof, said relevant features being defined within one or more classes of features, the method comprising the steps of: classifying said input data to generate a plurality of classes of similar features, and reducing the probability of the class or classes of features required to be identified within said input data and discarding input data having a low probability of containing a feature to be identified, to generate data of a reduced dimension; quantizing said reduced dimension data; generating an F1 score in respect of said quantized reduced dimension data and comparing said F1 score and probability percentages associated with the remaining classes in said reduced dimension data with a predefined threshold; and utilizing an adversarial neural network to generate data representative of one or more relevant features identified in said quantized reduced dimension data.

[0013] Other aspects of the invention provide a computer-implemented scene recognition system, a text reader system, a fall detection system and a voice recognition system, utilizing the method substantially as described above.

[0014] The method may be configured to be integrated into a wearable device, a mobile device, and other low memory and low compute devices.

[0015] The method may further comprise a continuous learning module configured to adapt to new data and environments. The method may comprise a preprocessing step in which preprocessing techniques optimized for low-compute devices are applied to said input data.

[0016] The preprocessing step may comprise one or more of noise reduction, contrast adjustment and normalization.

[0017] The invention also relates to a computer-implemented system for recognizing and processing input data, the system comprising a processor and memory in which is stored executable instructions configured, under control of the processor, to perform the method of any herein described, wherein said system comprises one or more of a scene recognition module, a text recognition module, and a sensor data processing module.

[0018] According to a further aspect there is a computer-implemented scene recognition system for a low memory and low compute device, the system comprising a processor and memory in which is stored executable instructions which, under control of the processor, are configured to execute a method comprising the steps of: o in an image acquisition module, capturing real-time video feeds containing scenes from various environments; o in an image preprocessing module, enhancing image quality for improved scene recognition accuracy; o in a scene recognition module, utilizing convolutional neural networks (CNNs) to identify and categorize scenes; o in a temporal analysis module, utilizing recurrent neural networks (RNNs) to analyze sequences of frames over time and understand context and movement within a scene; and o in an attention mechanism, focusing on the most relevant parts of the scene; and o generating output data representative of the most relevant parts of said scene; wherein the method further comprises utilizing an optimization method according to any herein described in the scene recognition module and / or the temporal analysis module to reduce the computational burden thereof.

[0019] The system may further comprise a continuous learning module to enable the system to adapt to new scenes and environments.

[0020] The image preprocessing module may apply noise reduction, binarization, contrast adjustment, and normalization using image processing techniques optimized for low compute devices.

[0021] The scene recognition module may include several convolutional layers followed by pooling layers and fully connected layers, and is trained on a large dataset of labeled images.

[0022] The temporal analysis module may use Long Short-Term Memory (LSTM) networks to process the temporal sequence of features provided by the CNN, enhancing the system's ability to recognize activities and dynamic environments.

[0023] The attention mechanism may weight the importance of different parts of the input image sequence, dynamically adjusting focus to improve recognition accuracy.

[0024] The continuous learning module may use incremental learning algorithms to update the model with new data without requiring a complete retraining.

[0025] The optimization techniques may include quantization to reduce the size of the neural network and pruning to remove unnecessary neurons and connections, ensuring fast and efficient operation on low memory and low compute devices.

[0026] According to a further aspect there is a computer-implemented text recognition and translation system for low memory and low compute devices, such as smart spectacles, the system comprising a processor and a memory in which is stored executable instructions which, under the control of the processor, are configured to perform a method comprising the steps of: o in an image acquisition module, capturing images or video frames containing text; o in an image preprocessing module, enhancing image quality for improved OCR accuracy; o in a text detection module, identifying and isolating text regions within the captured image; o in an OCR module, converting detected text regions into machine- readable text; o in an NLP module, processing the recognized text for translation and contextual understanding; o in a text-to-speech module, converting the processed text into audio; the method further comprises utilizing an optimization method according to any method herein described in the text detection module, the OCR module and / or the NLP module to reduce the computational burden thereof; the system further comprising a spectacle-based user interface configured to provide an interface for users to interact with the system.

[0027] The image preprocessing module may apply noise reduction, binarization, skew correction, and contrast adjustment using image processing techniques optimized for low compute devices.

[0028] The text detection module may uses a mobile-optimized version of the EAST text detector to identify text regions in real-time.

[0029] The OCR module may include a custom-trained neural network optimized for recognizing diverse fonts and languages on low memory and compute devices.

[0030] The NLP module may use lightweight Transformer-based models, such as DistilBERT or TinyBERT, for real-time text translation and contextual analysis. The text-to-speech module may utilize LiteTacotron and WaveLite to generate naturalsounding speech optimized for minimal memory and computational overhead.

[0031] According to a further aspect of the present inventio there is provide a computer- implemented fall detection system for low memory and low compute devices, such as smart spectacles, the system comprising an accelerometer sensor and a gyroscope sensor configured to capture real-time motion data from a user’s movements, the system further comprising a processor and a memory in which is stored executable instructions which, under control of the processor, are configured to perform a method comprising the steps of: o in a data preprocessing module, enhancing fall detection accuracy by filtering noise, normalizing data, and extracting relevant features; o in a fall detection module, utilizing lightweight machine learning algorithms to analyze preprocessed sensor data and distinguish between falls and other activities; o in an alert system, sending alerts with the user's location to predefined emergency contacts and automatically calling emergency services if the user does not respond within a set time; method further comprises utilizing an optimization method according to any method herein described in the fall detection module to reduce the computational burden thereof; the system further comprising a spectacle-based user interface configured to provide an interface for users to interact with the system.

[0032] The data preprocessing module may apply noise filtering, normalization, and feature extraction using signal processing techniques optimized for low compute devices.

[0033] The fall detection module may employ lightweight machine learning algorithms, such as decision trees, support vector machines (SVM), or a simplified neural network, trained to distinguish between falls and other activities. The alert system may use GPS data from the user's mobile device to determine the user's location and sends alerts to predefined emergency contacts.

[0034] The alert system may automatically call emergency services if the user does not respond within a set time, ensuring timely assistance.

[0035] According to a further aspect of the present invention there is provided a computer- implemented personal guide calling system for low memory and low compute devices, such as smart spectacles, the system comprising a processor and a memory in which is stored executable instructions which, under control of the processor, are configured to perform a method comprising the steps of: o in a voice recognition module, recognizing and processing voice commands from the user; o in an Al-based decision-making module, analyzing the voice commands and determine the appropriate action; o in a communication module, initiating calls, sending text messages, sharing the user's location information, and establishing a remote connection with a call center; o in a prioritization module, prioritizing contacts based on the urgency of the situation; o in a remote assistance module, allowing a call center employee to see through a camera located in the vicinity of a said low memory and low compute device, and assisting the user; wherein the method further comprises utilizing an optimization method according to any method hereon described in the decision-making module to reduce the computational burden thereof; the system further comprising a spectacle-based user interface configured to provide an interface for users to interact with the system. The voice recognition module may utilize lightweight algorithms optimized for low compute devices and includes noise reduction techniques to improve recognition accuracy.

[0036] The Al-based decision-making module may use natural language processing and decision tree algorithms to interpret commands and determine appropriate actions.

[0037] The communication module may integrate with the user's mobile device to use its communication capabilities for sending text messages, making voice calls, sharing GPS location data, and establishing a video connection with a call center employee.

[0038] The prioritization module may employ Al algorithms to dynamically prioritize contacts based on the urgency of the situation and user preferences.

[0039] The remote assistance module may establish a secure video connection with the call center, optimized for low bandwidth usage to ensure smooth video streaming on low memory and low compute devices.

[0040] These and other aspects of the invention will be apparent from the following Detailed Description.

[0041] Brief Description of the Drawings

[0042] Embodiments of the present invention will now be described, by way of examples only, and with reference to the accompanying drawings, in which:

[0043] Figure 1 is a schematic block diagram illustrating an example computer implemented system;

[0044] Figure 2 is a schematic diagram illustrating the principal features of a user device of an computer-implemented system;

[0045] Figure 3 is a schematic flow diagram illustrating the principal functions of an example computer-implemented scene recognition system; Figure 4 is a schematic flow diagram illustrating the principal functions of an example computer-implemented text reader;

[0046] Figure 5 is a schematic flow diagram illustrating the principal functions of an example computer-implemented fall detection system;

[0047] Figure 6 is a schematic flow diagram illustrating the principal functions of an example voice recognition system; and

[0048] Figure 7 is a schematic flow diagram illustrating the principal functions of an example optimization method suitable for use in the systems of Figures 3 to 6.

[0049] Detailed Description

[0050] Example embodiments of the invention utilize a unique combination of novel advanced deep learning algorithms and continuous learning techniques, specifically optimized for low memory and low compute devices. Such embodiments can be applied in a wide variety of fields including, but not limited to, healthcare (e.g. monitoring patient vitals), security (e.g. real-time surveillance analysis), transportation (e.g. autonomous vehicle navigation), and personal assistance (e.g. virtual personal assistants). The versatility and efficiency of example embodiments of the invention make it suitable for a wide range of applications, ensuring high accuracy and performance, even on resource- constrained devices.

[0051] Referring to Figure 1 of the drawings, a computer-implemented assistive technology system 100 is illustrated. The assistive technology system 100 comprises a communications server apparatus 102 and a user device 104. These devices 102, 104 are connected in the communication network 108 (for example, the internet) via communication links 110, 112 implementing, for example, internet communication protocols. The user device 104 may be able to communicate through other communications networks, such as public switched telephone networks (PSTN networks), including mobile cellular communication networks, but these are omitted from Figure 1 for the sake of clarity. Communications server apparatus 102 may be a single server, as illustrated schematically in Figure 1 , or have the functionality performed by the server apparatus 102 distributed across multiple server components. Indeed, in some example embodiments, at least some of the server functionality may be incorporated in the user device 104 itself, eliminating the need for the network 108 and communications links 110, 112.

[0052] In the example of Figure 1 , communications server apparatus 102 may comprise a number of individual components including, but not limited to, one or more microprocessors 116, a memory 118 (e.g. volatile memory such as RAM for the loading of executable instructions 120, the executable instructions 120 defining the functionality of the communications server apparatus 102, carried out under control of the processor 116. Communications server apparatus 102 also comprises an input / output module 122 (which may include a transmitter / receiver module) allowing the server to communicate over the communication network 108 (where applicable), receive data representative inputs (from the user device 104) and output a response.

[0053] Communications server apparatus 102 also comprises a database (and other data stores) 126, as well as document libraries 124. In this embodiment, the database (and data stores) 126 and document libraries 124 are part of the communications server apparatus 102, however, it should be appreciated that the database (and data stores) 126 and document libraries 124 can be separated from the communications server apparatus 102, and be connected thereto via the communication network 108 or via another communication link (not shown). Indeed, the entire communications server apparatus 102 could be implemented in the Cloud, rather than on a specific device or set of devices, and the invention is not intended to be limited in this regard.

[0054] The user device 104, incorporated in, for example, smart spectacles of the type referenced above, may comprise a number of individual components including, but not limited to, one or more microprocessors 128, a memory 130 (e.g. a volatile memory such as RAM) for the loading of executable instructions 132, the executable instructions (e.g. in the form of an app) defining the functionality of the user communications device 104, under control of the processor 128. The user device 104 also comprises an input / output module 134 (which may or may not include a transmitter / receiver module), allowing the user communications device 104 to communicate with the communications server apparatus 102 over the network 108 (if applicable and required).

[0055] Referring additionally to Figure 2 of the drawings, the executable instructions 132 may be configured to cause the user device to perform one or more of a plurality of different functions, including scene recognition 200, real-time text recognition and translation 300, fall detection 400, and personal guide calling and remote assistance 500.

[0056] The scene recognition ‘system’ 200 comprises an image acquisition module 210, an image preprocessing module 212 a scene recognition module 214 utilizing convolutional neural networks (CNNs), an RNN (recurrent neural network)-based temporal analysis module 216, an attention mechanism 218, a continuous learning module 220 and an optimization module utilizing optimization techniques such as quantization and pruning.

[0057] The real-time text recognition and translation ‘system’ 300 comprises an image acquisition module 312 and image preprocessing module 314, a text detection module 316, an OCR module 318, an NLP (natural language processing) module 320, a text-to- speech (TTS) module 322, and a user interface 138 (see also Figure 1 ).

[0058] The fall detection ‘system’ 400 comprises an accelerometer sensor 412, a gyroscope sensor 414, a preprocessing module 416, a machine learning - based fall detection module 418, an alert system 420, and a user interface 138 (see also Figure 1 ).

[0059] The personal guide and remote assistance ‘system’ 500 comprises a voice recognition module 512, an Al-based decision making module 514, a communication module 516, a prioritization module 518, a remote assistance module 520 and a user interface 138 (see also Figure 1 ).

[0060] All, or just one or more of the ‘systems’ 200, 300, 400, 500 may be integrated into the user device 104.

[0061] Scene recognition systems have become increasingly important in a variety of applications, particularly for providing contextual information and enhancing user experience in augmented reality (AR) and assistive technologies. Traditional scene recognition systems often struggle with accuracy and efficiency, especially when deployed on devices with limited computing resources. The example scene recognition system described hereinafter addresses these limitations by incorporating advanced deep learning algorithms and continuous learning techniques, specifically optimized for low memory and low compute devices such as, for example, smart spectacles, for example, of the type referenced above. Thus the following example scene recognition system 200 provides a real-time scene recognition system designed for low memory and low compute devices. The scene recognition system utilizes a combination of deep learning algorithms to analyze and interpret real-time video feeds captured by the camera of a mobile device. The system is designed to identify and categorize different environments, objects, and activities, providing contextual information an suggestions to the user. The innovative scene recognition system 200 uses advanced convolutional neural networks (CNNs) and recurrent neural networks (RNNs) and attention mechanisms, as well s continuous learning techniques to adapt to new scenes, and the system is specifically optimized for implementation in low memory and low compute devices, such as in spectacle-based devices with low memory and low compute capability. The system 200 captures real-time video feeds, preprocesses the images to enhance recognition accuracy, identifies and categorizes scenes, and provides contextual information and suggestions to the user. The system 200 employs deep learning algorithms, continuous learning, and optimization techniques to provide a robust and efficient solution for real-time scene recognition.

[0062] Referring to Figure 3 of the drawings, the real-time scene recognition system 200 is illustrated schematically in terms of its main functions.

[0063] 1. Image Acquisition (230):

[0064] • Component: Camera Module

[0065] Function: Captures real-time video feeds containing scenes from various environments. • Technical Detail: High-resolution miniature camera with autofocus to ensure clear capture of scenes under different lighting conditions, optimized for the limited space in spectacle frames.

[0066] 2. Preprocessing (232):

[0067] • Component: Image Preprocessing Module

[0068] • Function: Enhances image quality to improve scene recognition accuracy.

[0069] • Technical Detail: Includes steps such as noise reduction, binarization, contrast adjustment, and normalization using image processing techniques optimized to run efficiently on low compute devices.

[0070] 3. Scene Recognition Module (234):

[0071] • Component: Convolutional Neural Networks (CNNs)

[0072] • Function: Handles image data to identify and categorize scenes.

[0073] • Technical Detail: o Architecture: Includes several convolutional layers followed by pooling layers and fully connected layers. o Feature Extraction: Convolutional layers automatically learn spatial hierarchies of features from input images. Pooling layers reduce the dimensionality, retaining only the most critical features, and fully connected layers aggregate these features to make predictions. o Training: The CNN is trained on a large dataset of labeled images, enabling it to recognize a wide variety of scenes. The architecture is optimized for memory and computational efficiency.

[0074] The ResNet architecture, which is the backbone of many modem CNNs used in scene recognition, is introduced in “Deep Residual Learning for Image Recognition”, by Kaiming He, et al

[0075] 4. Temporal Analysis Module (236): • Component: Recurrent Neural Networks (RNNs)

[0076] • Function: Analyzes sequences of frames over time to understand context and movement within a scene.

[0077] • Technical Detail: o Temporal Analysis: Uses Long Short-Term Memory (LSTM) networks to process the temporal sequence of features provided by the CNN, enhancing the system's ability to recognize activities and dynamic environments. o Integration: The output from the CNN is fed into the RNN, which processes the sequence to provide a comprehensive understanding of the scene.

[0078] The use of temporal analysis of video sequences is discussed in, for example, “Long Short-Term Memory”, by Sepp Hochreiter and Jurgen Schmidhuber

[0079] 5. Attention Mechanism (238):

[0080] • Component: Attention Mechanism. Suitable attention mechanisms are discussed in “Show, Attend and Tell: Neural Image Caption Generation with Visual Attention”, by Kelvin Xu, et al.

[0081] • Function: Focuses on the most relevant parts of the scene.

[0082] • Technical Detail: o Focus: Weights the importance of different parts of the input image sequence, dynamically adjusting focus to improve recognition accuracy. o Implementation: Uses attention mechanisms to allow the system to focus on relevant objects and activities, particularly useful in complex environments where multiple objects and activities are present.

[0083] 6. Continuous Learning (240):

[0084] • Component: Continuous Learning Module • Function: Enables the system to adapt to new scenes and environments.

[0085] • Technical Detail: o Adaptability: Achieved through incremental learning algorithms that update the model with new data without requiring a complete retraining. o User Feedback: Incorporates user interactions and feedback into the learning process, allowing the system to refine its accuracy based on real- world usage.

[0086] Incremental learning techniques are discussed in “Incremental Learning of Object Detectors without Catastrophic Forgetting”, by Shmelkov.

[0087] 7. Optimization Techniques (242):

[0088] • Component: Optimization Module

[0089] • Function: Ensures the system runs efficiently on low memory and low compute devices.

[0090] • Technical Detail: o Quantization: Reduces the size of the neural network without significantly compromising accuracy. o Pruning: Removes unnecessary neurons and connections in the network, further optimizing the model for faster inference.

[0091] Techniques for the optimization of neural networks for deployment on mobile devices is outlined in “Quantization and Training of Neural Networks for Efficient Integer- Arithmetic-Only Inference”, by Jacob, et al.

[0092] Unique Features:

[0093] 1. Adaptive Learning: o The system's ability to continuously learn and adapt to new scenes ensures that recognition accuracy improves over time, making it more reliable for users. 2. Real-Time Processing: o The integration of CNNs and RNNs allows for real-time processing of video feeds, providing instant feedback and contextual information to users.

[0094] 3. User-Centric Design: o The system leverages user feedback to enhance its learning process, ensuring that it remains relevant and accurate in diverse environments.

[0095] 4. Efficient Deployment: o Optimization techniques like quantization and pruning make the system suitable for deployment on low memory and low compute devices, ensuring fast and efficient operation without heavy computational requirements.

[0096] Text recognition and translation systems are increasingly important in a variety of applications, particularly for visually impaired users. Traditional OCR systems often struggle with accuracy in challenging conditions such as low light, distorted text, and diverse fonts. Furthermore, the integration of translation and contextual understanding capabilities is limited in existing solutions. The text recognition and translation system 300 addresses these limitations by incorporating advanced deep learning algorithms and continuous learning techniques, specifically optimized for low memory and low compute devices, such as smart spectacles, for example, of the type referenced above. Accordingly, the text recognition system 300 utilizes optical character recognition (OCR) technology, enhanced with natural language processing (NLP) to detect and translate text in real time from various surfaces and contexts. The system 300 is designed to recognize text from books, signs and digital screens and convert it into audio, making it accessible for visually impaired users. The system 300 is specifically optimized for implementation in, for example, spectacle-based devices with low memory and low compute capabilities. The system 300 captures images or video frames containing text, preprocesses the images to enhance OCR accuracy, detects and recognizes text regions, processes the text using NLP for translation and contextual understanding, and converts the text into audio. The system employs deep learning algorithms, continuous learning, and optimization techniques to provide a robust and efficient solution for realtime text recognition and translation.

[0097] Referring to Figure 4 of the drawings, the real-time text recognition and translation system 300 is illustrated schematically in terms of its main functions.

[0098] 1. Image Acquisition (330):

[0099] • Component: Camera Module

[0100] • Function: Captures images or video frames containing text from various sources.

[0101] • Technical Detail: High-resolution miniature camera with autofocus to ensure clear capture of text under different lighting conditions, optimized for the limited space in spectacle frames.

[0102] 2. Preprocessing (332):

[0103] • Component: Image Preprocessing Module

[0104] • Function: Enhances image quality to improve OCR accuracy.

[0105] • Technical Detail: Includes steps such as noise reduction, binarization, skew correction, and contrast adjustment using image processing techniques like Gaussian blur, thresholding, and Hough transform. These processes are optimized to run efficiently on low compute devices.

[0106] 3. Text Detection (334):

[0107] • Component: Text Detection Module

[0108] • Function: Identifies and isolates text regions within the captured image.

[0109] • Technical Detail: Uses lightweight deep learning-based algorithms, such as a mobile-optimized version of the EAST (Efficient and Accurate Scene Text) detector, to detect text regions with high precision. The algorithm outputs bounding boxes around detected text, tailored to the computational limits of spectacle-based devices. ptical Character Recognition (OCR) (336):

[0110] • Component: OCR Module

[0111] • Function: Converts detected text regions into machine-readable text.

[0112] • Technical Detail: Implemented using Tesseract OCR enhanced with deep learning techniques for improved accuracy. The OCR module includes a custom- trained neural network to recognize a wide variety of fonts and languages, optimized for memory and computational efficiency. atural Language Processing (NLP) (338):

[0113] • Component: NLP Module

[0114] • Function: Processes the recognized text for translation and context understanding.

[0115] • Technical Detail: o Text Normalization: Corrects OCR errors, handles abbreviations, and standardizes text. o Translation: Uses lightweight Transformer-based models (e.g., DistilBERT, TinyBERT) for real-time text translation into multiple languages, optimized for low memory usage. o Contextual Analysis: Applies NLP techniques for semantic understanding, such as named entity recognition (NER), part-of-speech tagging, and sentiment analysis, all tailored for low compute environments. ext-to-Speech (TTS) (340):

[0116] • Component: TTS Module

[0117] • Function: Converts the processed text into audio.

[0118] Technical Detail: Utilizes state-of-the-art TTS engines (e.g., LiteTacotron,

[0119] WaveLite) to generate natural-sounding speech, optimized for minimal memory and computational overhead. The system allows for customization of voice and intonation.

[0120] 7. User Interface (138):

[0121] • Component: Spectacle-Based User Interface

[0122] • Function: Provides an interface for users to interact with the system.

[0123] • Technical Detail: Developed using lightweight frameworks to ensure smooth operation on low memory and low compute devices. The interface includes features like real-time text reading, language selection, and voice customization, accessible through simple gestures or voice commands.

[0124] Unique Features:

[0125] Enhanced OCR Accuracy: o Deep Learning Integration: The integration of deep learning techniques with traditional OCR methods significantly enhances text recognition accuracy, especially in challenging conditions such as low light or distorted text. o Custom Training: The OCR module is trained on a diverse dataset that includes various fonts, sizes, and languages, improving its versatility and performance while being optimized for low memory and low compute devices.

[0126] Real-Time Processing: o Efficient Algorithms: The use of the mobile-optimized EAST text detector and lightweight Transformer-based models enables real-time text detection, recognition, and translation, providing immediate feedback to users. o Optimized Pipeline: The entire processing pipeline is optimized for speed and efficiency, ensuring that users experience minimal latency even on low compute devices like smart spectacles. Multi-Language Support: o NLP Integration: The system's ability to translate text into multiple languages using lightweight NLP models makes it accessible to a global audience. o Contextual Understanding: The NLP module's contextual analysis capabilities ensure accurate translations and provide meaningful insights into the recognized text.

[0127] Accessibility Features: o TTS Customization: The text-to-speech module offers customizable voice options, allowing users to select their preferred voice and intonation for audio output, optimized for low memory usage. o User-Friendly Interface: The spectacle-based user interface is designed with accessibility in mind, featuring intuitive controls and support for voice commands, ensuring ease of use for visually impaired users.

[0128] Technical Implementation Details:

[0129] 1. Preprocessing Module: o Noise Reduction: Gaussian blur is applied to smooth the image and reduce high-frequency noise. o Binarization: Adaptive thresholding converts the image to a binary format, making text regions more distinct. o Skew Correction: Hough transform is used to detect and correct skewed text lines, ensuring accurate OCR results.

[0130] 2. Text Detection Module: o Mobile-Optimized EAST Detector: A pre-trained EAST model, optimized for mobile devices, is used to identify text regions in real-time. The model is fine-tuned on a custom dataset to improve detection accuracy for various text styles and languages. 3. OCR Module: o Tesseract OCR: Enhanced with a custom-trained neural network for recognizing diverse fonts and languages. The network includes multiple layers of convolutional and recurrent units to handle complex text patterns, optimized for low memory and compute environments.

[0131] 4. NLP Module: o Text Normalization: A sequence-to-sequence model corrects common OCR errors and normalizes text. o Translation: A lightweight Transformer-based model, such as DistilBERT or TinyBERT, is used for text translation. The model is fine-tuned for realtime performance with minimal memory usage. o Contextual Analysis: Techniques like NER and POS tagging provide additional context for accurate translation and understanding.

[0132] 5. TTS Module: o LiteTacotron: Converts the processed text into a mel-spectrogram, which is then synthesized into speech using a WaveLite vocoder. The TTS system supports multiple languages and voice options, optimized for low memory and computational overhead.

[0133] Thus, the text recognition and translation system 300 utilizes a unique combination of advanced components and techniques to provide a novel and powerful solution for realtime text recognition and translation, making it accessible for a wide range of users, especially those with visual impairments, while being optimized for low memory and low compute devices such as smart spectacles, for example, of the type referenced above.

[0134] Fall detection systems are crucial for ensuring the safety and wellbeing of individuals, particularly the elderly and those with mobility issues. Traditional fall detection systems often suffer from high false positive rates and require significant computational resources, limiting their effectiveness in low memory and low compute wearable devices. Accordingly, the fall detection system 400 seeks to address these limitations by incorporating advanced machine learning algorithms optimized for low memory and low compute devices such as smart spectacles, for example, of the type referenced above.

[0135] The fall detection system 400 comprises accelerometer and gyroscope sensors integrated into a spectacle-based wearable device to accurately detect falls. The system 400 employs machine learning algorithms to distinguish between falls and other activities, significantly reducing false positives. Upon detecting a fall, the system may be configured to send an alert with the user’s location to emergency contacts and can automatically call the emergency services if the user does not respond within a set time. The system 400 is specifically optimized for low memory and low compute devices such as smart spectacles, for example, of the type referenced above. The system 400 comprises an accelerometer sensor, a gyroscope sensor, a preprocessing module, a machine learning - based fall detection module, an alert system and a user interface. The system 400 captures and processes sensor data to detect falls accurately, distinguishes between falls and other activities using machine learning algorithms, and sends alerts with the user’s location to emergency contacts. The system 400 is optimized for low power and low compute wearable devices, ensuring efficient operation and timely response.

[0136] Referring to Figure 5 of the drawings, the fall detection system 400 is illustrated schematically in terms of its main functions.

[0137] 1. Sensor Data Acquisition (430):

[0138] • Component: Accelerometer and Gyroscope Sensors. Various techniques for fall detection using accelerometer sensors are reviewed in “A Comprehensive Review on Accelerometer-Based Fall Detection”, by S.K. Nambiaret al.

[0139] • Function: Capture real-time motion data from the user's movements.

[0140] • Technical Detail: High-precision accelerometer and gyroscope sensors integrated into the spectacle frame, capturing three-axis acceleration and rotational motion data. reprocessing (432):

[0141] • Component: Data Preprocessing Module

[0142] • Function: Preprocesses raw sensor data to enhance fall detection accuracy.

[0143] • Technical Detail: Includes steps such as noise filtering, normalization, and feature extraction using signal processing techniques optimized for low compute devices.all Detection Module (434):

[0144] • Component: Machine Learning-Based Fall Detection

[0145] • Function: Analyzes preprocessed sensor data to detect falls.

[0146] • Technical Detail: o Feature Extraction: Extracts relevant features from the sensor data, such as peak acceleration, angular velocity, and impact duration. o Algorithm: Employs lightweight machine learning algorithms, such as decision trees, support vector machines (SVM), or a simplified neural network, trained to distinguish between falls and other activities. The model is optimized for low memory usage and computational efficiency.lert System (436):

[0147] • Component: Alert and Notification Module

[0148] • Function: Sends alerts and notifications upon detecting a fall.

[0149] • Technical Detail: o Location Tracking: Uses GPS data from the user's mobile device to determine the user's location. o Emergency Contacts: Sends an alert with the user's location to predefined emergency contacts. o Automatic Emergency Call: Automatically calls emergency services if the user does not respond within a set time, ensuring timely assistance. 5. User Interface (138):

[0150] • Component: Spectacle-Based User Interface

[0151] • Function: Provides an interface for users to interact with the system.

[0152] • Technical Detail: Developed using lightweight frameworks to ensure smooth operation on low memory and low compute devices. The interface includes features like setting emergency contacts, adjusting sensitivity levels, and viewing fall history.

[0153] Unique Features:

[0154] Accurate Fall Detection: o Machine Learning Integration: The integration of machine learning algorithms significantly enhances fall detection accuracy, reducing false positives by distinguishing between falls and other activities. Different machine learning algorithms for fall detection using wearable sensors are evaluated in “Real-Time Fall Detection with Wearable Sensors: Evaluating Machine Learning Algorithms and Performance Trade-Offs”, by M.C.C. Pimentel, et al. o Optimized for Wearables: The algorithms are specifically optimized for low memory and low compute devices, ensuring efficient operation on smart spectacles. Optimization of deep learning models for deployment on mobile and edge devices is discussed in “Lightweight Deep Learning Models for Mobile and Edge Devices: Progress, Challenges, and Opportunities”, by Y. Cheng, et al.

[0155] Real-Time Processing: o Efficient Algorithms: The use of lightweight machine learning models enables real-time processing of sensor data, providing immediate detection and response. o Optimized Pipeline: The entire processing pipeline is optimized for speed and efficiency, ensuring minimal latency in fall detection and alerting.

[0156] User-Centric Design: o Customizable Alerts: The system allows users to customize emergency contacts and alert settings, ensuring that the alerts are sent to the appropriate recipients. o Intuitive Interface: The spectacle-based user interface is designed with ease of use in mind, featuring intuitive controls and support for voice commands.

[0157] Reliable Operation: o Continuous Monitoring: The system continuously monitors the user's movements, providing reliable fall detection and timely alerts. o Automatic Emergency Response: The automatic emergency call feature ensures that help is summoned even if the user is unable to respond, enhancing user safety.

[0158] Technical Implementation Details:

[0159] Preprocessing Module: o Noise Filtering: Applies low-pass filters to smooth the sensor data and reduce high-frequency noise. o Normalization: Standardizes the sensor data to a common scale, facilitating accurate feature extraction. o Feature Extraction: Identifies key features from the sensor data, such as peak acceleration, angular velocity, and impact duration, which are critical for fall detection.

[0160] Fall Detection Module: o Lightweight Machine Learning Algorithms: Uses decision trees, support vector machines (SVM), or a simplified neural network trained on a dataset of labeled fall and non-fall events. The model is fine-tuned to achieve a balance between accuracy and computational efficiency. o Training: The model is trained on a large dataset of sensor data, including both fall and non-fall events, to accurately distinguish between different types of activities.

[0161] Alert System: o Location Tracking: Integrates GPS data from the user's mobile device to determine the user's location accurately. o Emergency Contacts: Allows users to set predefined emergency contacts who will receive alerts in case of a fall. o Automatic Emergency Call: Initiates an automatic call to emergency services if the user does not respond within a set time, ensuring timely assistance.

[0162] By integrating these advanced components and techniques, the fall detection system 400 provides a unique and powerful solution for real-time fall detection on low memory and low compute devices such as smart spectacles, for example of the type referenced above, enhancing user safety and ensuring timely assistance in case of falls.

[0163] Personal assistance systems are crucial for providing timely help and enhancing the safety and convenience of users, particularly those with disabilities or in emergency situations. Traditional systems often require significant computational resources, limiting their effectiveness in low memory and low compute wearable devices. The personal guide calling system 500 with remote assistance aims to address these limitations by incorporating advanced voice recognition and Al techniques, specifically optimized for low memory and low compute devices such as smart spectacles, for example, of the type referenced above. Additionally, it provides a unique remote assistance feature, whereby call centre employees can assist users by viewing their environment through a camera on the wearable device. Accordingly, the system 500 provides real-time voice-activated personal assistance and remote navigation assistance using low memory and low compute devices, such as smart spectacles. The personal guide calling system utilizes voice recognition and Al to connect users with a predefined list of contacts for assistance. The system can recognize voice commands to initiate calls, send text messages, or share the user’s location. It may include a feature to remotely connect with a call centre employee who can see through the camera on the wearable device and visually impaired users to navigate. The system 500 prioritises contacts based on the urgency of the situation and provides real-time updates to the user and their contacts.

[0164] Accordingly, the system 500 is specifically optimized for implementation in wearable devices such as smart spectacles, having low memory and low compute capabilities. It is configured to recognize voice commands to initiate calls, send text messages, or share the user’s location, and is able to prioritise contacts based on the urgency of the situation, and provide real-time updates to the user and their contacts. It may also include a feature whereby a call centre employee can view the user’s environment through a camera on the wearable device and provide navigation assistance. The system 500 employs advanced Al algorithms and optimization techniques to ensure efficient operation on low memory and low compute wearable devices.

[0165] Referring to Figure 6 of the drawings, the personal guide calling system 500 is illustrated schematically in terms of its main functions.

[0166] 1. Voice Recognition (530):

[0167] • Component: Voice Recognition Module

[0168] • Function: Recognizes and processes voice commands from the user.

[0169] • Technical Detail: Utilizes lightweight voice recognition algorithms optimized for low compute devices. The module can accurately recognize a predefined set of commands, even in noisy environments.

[0170] 2. Decision-Making (532): Component: Al-Based Decision-Making Module

[0171] • Function: Analyzes the voice commands and determines the appropriate action.

[0172] • Technical Detail: Uses natural language processing (NLP) and decision tree algorithms to interpret commands and decide whether to initiate a call, send a text message, share the user's location, or connect with a call center employee.

[0173] 3. Communication (534):

[0174] • Component: Communication Module

[0175] • Function: Handles the initiation of calls, sending of text messages, sharing of location information, and establishing a remote connection with a call center.

[0176] • Technical Detail: Integrates with the user's mobile device to use its communication capabilities. The module can send text messages, make voice calls, share GPS location data, and establish a video connection with a call center employee.

[0177] 4. Prioritization (536):

[0178] • Component: Prioritization Module

[0179] • Function: Prioritizes contacts based on the urgency of the situation.

[0180] • Technical Detail: Uses Al algorithms to assess the urgency of the situation and prioritize contacts accordingly. The system can dynamically adjust the priority list based on the context and user preferences.

[0181] 5. Remote Assistance (538):

[0182] • Component: Remote Assistance Module

[0183] • Function: Allows a call center employee to see through the spectacle's camera and assist the user.

[0184] • Technical Detail: Establishes a secure video connection with the call center, enabling the employee to view the user's environment in real-time. The module is optimized for low bandwidth usage to ensure smooth video streaming on low memory and low compute devices.

[0185] 6. User Interface (138):

[0186] • Component: Spectacle-Based User Interface

[0187] • Function: Provides an interface for users to interact with the system.

[0188] • Technical Detail: Developed using lightweight frameworks to ensure smooth operation on low memory and low compute devices. The interface includes voice command activation, visual indicators for system status, emergency contact management, and remote assistance activation.

[0189] Unique Features:

[0190] Accurate Voice Recognition: o Optimized Algorithms: The use of lightweight voice recognition algorithms ensures accurate recognition of commands, even on low compute devices. o Noise Reduction: The system includes noise reduction techniques to improve recognition accuracy in various environments.

[0191] Intelligent Decision-Making: o NLP Integration: The Al-based decision-making module leverages NLP to understand and interpret user commands accurately. o Contextual Understanding: The system can assess the context and urgency of the situation to make appropriate decisions.

[0192] Efficient Communication: o Seamless Integration: The communication module integrates seamlessly with the user's mobile device, enabling reliable calls, messages, and location sharing. o Real-Time Updates: The system provides real-time updates to the user and their contacts, ensuring timely assistance.

[0193] Dynamic Prioritization: o Al-Based Prioritization: The system dynamically prioritizes contacts based on the urgency of the situation, ensuring that the most appropriate contacts are notified first. o User Preferences: The prioritization module can be customized based on user preferences and past interactions.

[0194] Remote Assistance: o Real-Time Navigation Help: The remote assistance module allows a call center employee to view the user's environment through the spectacle's camera and provide real-time navigation help. o Low Bandwidth Optimization: The video connection is optimized for low bandwidth usage, ensuring smooth and reliable video streaming on low memory and low compute devices.

[0195] Technical Implementation Details:

[0196] 1. Voice Recognition Module: o Lightweight Algorithms: Utilizes efficient voice recognition algorithms such as PocketSphinx, optimized for low memory and compute devices. The use of efficient algorithms for voice recognition is discussed in “End-to-End Speech Recognition in English and Mandarin”, by J. Li, et al. o Noise Reduction: Applies noise reduction techniques to enhance voice command recognition in various environments. o End-to-end automatic speech recognition techniques are explored in “An Overview of End-to-End Automatic Speech Recognition”, by G. Saon, et al.

[0197] 2. Decision-Making Module: o NLP and Decision Trees: Combines natural language processing with decision tree algorithms to interpret and act on user commands. o Contextual Analysis: Uses contextual cues to understand the urgency of the command and make appropriate decisions. o Insights into efficient text-to-speech synthesis methods that can be integrated into the decision making module are discussed in “A Comprehensive Survey on Text-to-Speech Synthesis, byX. Tan, et al.

[0198] 3. Communication Module: o Mobile Integration: Integrates with the communication capabilities of the user's mobile device, including text messaging, voice calls, and GPS location sharing. o Efficient Protocols: Uses efficient communication protocols to ensure timely and reliable message delivery and video streaming.

[0199] 4. Prioritization Module: o Al Algorithms: Employs Al algorithms to dynamically prioritize emergency contacts based on the context and urgency of the situation. o User Preferences: Allows customization of the priority list based on user preferences and past interactions.

[0200] 5. Remote Assistance Module: o Secure Video Connection: Establishes a secure video connection with the call center, enabling real-time assistance. o Low Bandwidth Optimization: Uses compression and other optimization techniques to ensure smooth video streaming on low bandwidth and low compute devices. By integrating advanced components and techniques, the personal guide calling system 500 provides a unique and powerful solution for real-time voice-activated personal assistance and remote navigation help on low memory and low compute devices like smart spectacles, for example, of the type referenced above, ensuring timely help and enhancing user safety and convenience.

[0201] Thus, in the case of each of the innovative ‘systems’ 200, 300, 400, 500 described above, one of the key innovative features resides in the manner of processing the data in order to be able to implement deep learning algorithms and continuous learning, and achieve the desired accuracy in real time, in a device that has a small hardware footprint (i.e. low computational resources).

[0202] Irrespective of the nature of the input data (whether it be text data, voice data, image data or video data), the use of a unique optimization technique is utilized in order to significantly reduce the computational burden without loss of important data, thereby achieving, in real-time, an output of the desired high accuracy in a low memory low compute device.

[0203] The novel optimization technique is implemented with the use of a classification model that reduces the probability of the class required to be identified (known herein as dimensionality reduction). In the thus reduced dimension space, quantization methods are applied so that memory and processing requirement is reduced. The result would, of course, be an approximation, but as the number of classes is relatively low, and the variations between them are larger (in terms of features), this method works well and delivers sufficiently high accuracy. Following the classification and quantization steps, decision-making steps are performed, wherein an F1 score and probability percentages are checked against a predefined threshold and, subsequently, a ‘decision’ is made by two different ‘agents’ competing with each other to resolve the problem in the fastest possible time (Adversarial Neural Network).

[0204] Thus, and referring to Figure 7 of the drawings, in an example method, that can be used in any of the ‘systems’ 200, 300, 400, 500 described above, for processing incoming data and generating, in real time, an output of high accuracy in a device having a small hardware footprint, the following principal steps are employed in respect of the input data 700:

[0205] Unified Processing Algorithm (710): o Despite varying input types (text, video, images), a single processing algorithm is used, demonstrating versatility and scalability.

[0206] Dimensionality Reduction via Classification (720): o Using classification for dimensionality reduction is a novel approach, simplifying the data and reducing the computational burden in subsequent steps.

[0207] Efficient Quantization (730): o Applying quantization in a lower-dimensional space ensures efficient processing on low-memory devices without significant loss in accuracy.

[0208] Competitive Decision-Making (740): o Implementing an adversarial neural network introduces a competitive element in decision-making, improving the speed and robustness of the system.

[0209] Tailored Thresholds (750): o Setting specific thresholds for F1 score and recognition percentages ensures a balance between precision and recall, catering to the unique requirements of each application.

[0210] In summary, the approach combines classification, quantization, and adversarial decision-making to achieve high accuracy on devices with small hardware footprints. The innovations in dimensionality reduction, efficient processing, and competitive decision-making make this method particularly effective and unique.

[0211] It will be apparent to a person skilled in the art from the foregoing description, that modifications and variations can be made to the described embodiments without departing from the scope of the invention as defined in the appended claims.

Claims

CLAIMS1 . A method for optimizing input data and determining relevant features thereof, said relevant features being defined within one or more classes of features, the method comprising the steps of: classifying said input data to generate a plurality of classes of similar features, and reducing the probability of the class or classes of features required to be identified within said input data and discarding input data having a low probability of containing a feature to be identified, to generate data of a reduced dimension; quantizing said reduced dimension data; generating an F1 score in respect of said quantized reduced dimension media data and comparing said F1 score and probability percentages associated with the remaining classes in said reduced dimension data with a predefined threshold; and utilizing an adversarial neural network to generate data representative of one or more relevant features identified in said quantized reduced dimension data.

2. The method according to claim 1 , configured to be integrated into a wearable device, a mobile device, and other low memory and low compute devices.

3. A method according to claim 1 or claim 2, further comprising a continuous learning module configured to adapt to new data and environments.

4. A method according to any of claims 1 to 3, comprising a preprocessing step in which preprocessing techniques optimized for low-compute devices are applied to said input data.

5. A method according to claim 4, wherein said preprocessing step comprises one or more of noise reduction, contrast adjustment and normalization.

6. A computer-implemented system for recognizing and processing input data, the system comprising a processor and memory in which is stored executable instructions configured, under control of the processor, to perform the method of any of claims 1 to 5, wherein said system comprises one or more of a scenerecognition module, a text recognition module, and a sensor data processing module.

7. A computer-implemented scene recognition system for a low memory and low compute device, the system comprising a processor and memory in which is stored executable instructions which, under control of the processor, are configured to execute a method comprising the steps of: o in an image acquisition module, capturing real-time video feeds containing scenes from various environments; o in an image preprocessing module, enhancing image quality for improved scene recognition accuracy; o in a scene recognition module, utilizing convolutional neural networks (CNNs) to identify and categorize scenes; o in a temporal analysis module, utilizing recurrent neural networks (RNNs) to analyze sequences of frames over time and understand context and movement within a scene; and o in an attention mechanism, focusing on the most relevant parts of the scene; and o generating output data representative of the most relevant parts of said scene; wherein the method further comprises utilizing an optimization method according to any of claims 1 to 5 in the scene recognition module and / or the temporal analysis module to reduce the computational burden thereof.

8. The system of claim 7, further comprising a continuous learning module to enable the system to adapt to new scenes and environments.

9. The system of claim 7 or claim 8, wherein the image preprocessing module applies noise reduction, binarization, contrast adjustment, and normalization using image processing techniques optimized for low compute devices.

10. The system of any of claims 7 to 9, wherein the scene recognition module includes several convolutional layers followed by pooling layers and fully connected layers, and is trained on a large dataset of labeled images.11 . The system of any of claims 7 to 10, wherein the temporal analysis module uses Long Short-Term Memory (LSTM) networks to process the temporal sequence of features provided by the CNN, enhancing the system's ability to recognize activities and dynamic environments.

12. The system of any of claims 7 to 11 , wherein the attention mechanism weights the importance of different parts of the input image sequence, dynamically adjusting focus to improve recognition accuracy.

13. The system of any of claims 7 to 12, wherein the continuous learning module uses incremental learning algorithms to update the model with new data without requiring a complete retraining.

14. The system of any of claims 7 to 13, wherein the optimization techniques include quantization to reduce the size of the neural network and pruning to remove unnecessary neurons and connections, ensuring fast and efficient operation on low memory and low compute devices.

15. A computer-implemented text recognition and translation system for low memory and low compute devices, such as smart spectacles, the system comprising a processor and a memory in which is stored executable instructions which, under the control of the processor, are configured to perform a method comprising the steps of: o in an image acquisition module, capturing images or video frames containing text; o in an image preprocessing module, enhancing image quality for improved OCR accuracy; o in a text detection module, identifying and isolating text regions within the captured image;o in an OCR module, converting detected text regions into machine- readable text; o in an NLP module, processing the recognized text for translation and contextual understanding; o in a text-to-speech module, converting the processed text into audio; the method further comprises utilizing an optimization method according to any of claims 1 to 5 in the text detection module, the OCR module and / or the NLP module to reduce the computational burden thereof; the system further comprising a spectacle-based user interface configured to provide an interface for users to interact with the system;16. The system of claim 15, wherein the image preprocessing module applies noise reduction, binarization, skew correction, and contrast adjustment using image processing techniques optimized for low compute devices.

17. The system of claim 15 or claim 16, wherein the text detection module uses a mobile-optimized version of the EAST text detector to identify text regions in realtime.

18. The system of any of claims 15 to 17, wherein the OCR module includes a custom-trained neural network optimized for recognizing diverse fonts and languages on low memory and compute devices.

19. The system of any of claims 15 to 18, wherein the NLP module uses lightweight Transformer-based models, such as DistilBERT or TinyBERT, for real-time text translation and contextual analysis.

20. The system of any of claims 15 to 19, wherein the text-to-speech module utilizes LiteTacotron and WaveLite to generate natural-sounding speech optimized for minimal memory and computational overhead.21 . A computer-implemented fall detection system for low memory and low compute devices, such as smart spectacles, the system comprising an accelerometer sensor and a gyroscope sensor configured to capture real-time motion data from a user’s movements, the system further comprising a processor and a memory inwhich is stored executable instructions which, under control of the processor, are configured to perform a method comprising the steps of: o in a data preprocessing module, enhancing fall detection accuracy by filtering noise, normalizing data, and extracting relevant features; o in a fall detection module, utilizing lightweight machine learning algorithms to analyze preprocessed sensor data and distinguish between falls and other activities; o in an alert system, sending alerts with the user's location to predefined emergency contacts and automatically calling emergency services if the user does not respond within a set time; method further comprises utilizing an optimization method according to any of claims 1 to 5 in the fall detection module to reduce the computational burden thereof; the system further comprising a spectacle-based user interface configured to provide an interface for users to interact with the system.

22. The system of claim 21 , wherein the data preprocessing module applies noise filtering, normalization, and feature extraction using signal processing techniques optimized for low compute devices.

23. The system of claim 21 or claim 22, wherein the fall detection module employs lightweight machine learning algorithms, such as decision trees, support vector machines (SVM), or a simplified neural network, trained to distinguish between falls and other activities.

24. The system of any of claims 21 to 23, wherein the alert system uses GPS data from the user's mobile device to determine the user's location and sends alerts to predefined emergency contacts.

25. The system of any of claims 21 to 24, wherein the alert system automatically calls emergency services if the user does not respond within a set time, ensuring timely assistance.

26. A computer-implemented personal guide calling system for low memory and low compute devices, such as smart spectacles, the system comprising a processorand a memory in which is stored executable instructions which, under control of the processor, are configured to perform a method comprising the steps of: o in a voice recognition module, recognizing and processing voice commands from the user; o in an Al-based decision-making module, analyzing the voice commands and determine the appropriate action; o in a communication module, initiating calls, sending text messages, sharing the user's location information, and establishing a remote connection with a call center; o in a prioritization module, prioritizing contacts based on the urgency of the situation; o in a remote assistance module, allowing a call center employee to see through a camera located in the vicinity of a said low memory and low compute device, and assisting the user; wherein the method further comprises utilizing an optimization method according to any of claims 1 to 5 in the decision-making module to reduce the computational burden thereof; the system further comprising a spectacle-based user interface configured to provide an interface for users to interact with the system.

27. The system of claim 26, wherein the voice recognition module utilizes lightweight algorithms optimized for low compute devices and includes noise reduction techniques to improve recognition accuracy.

28. The system of claim 26 or claim 27, wherein the Al-based decision-making module uses natural language processing and decision tree algorithms to interpret commands and determine appropriate actions.

29. The system of any of claims 26 to 28, wherein the communication module integrates with the user's mobile device to use its communication capabilities for sending text messages, making voice calls, sharing GPS location data, and establishing a video connection with a call center employee.

30. The system of any of claims 26 to 29, wherein the prioritization module employs Al algorithms to dynamically prioritize contacts based on the urgency of the situation and user preferences.31 . The system of any of claims 26 to 30, wherein the remote assistance module establishes a secure video connection with the call center, optimized for low bandwidth usage to ensure smooth video streaming on low memory and low compute devices.

Images