Edge-executed machine learning for motion-responsive wearable radars
View Patent ↗An edge-executed machine learning approach for motion-responsive radar in wearables enables detection of approaching objects reliably in the presence of movement and vibrations of the user. The wearable early detection system can use a machine learning model using neural networks (such as convolutional or recurrent neural networks) to combine the information from various sensors, such as a radar and inertial sensors, and provide accurate identification of approaching objects under different states of motion and generate various kinds of alarms including audio, visual, vibrations, and haptic to the user and approaching vehicles directly or via a secondary wirelessly connected device.
1 . A wearable system adapted to detect an approaching object, comprising:
a radio frequency (RF) unit comprising:
a local RF source generating an output;
an RF transmitter comprising at least one antenna element driven in synchronization with the local RF source;
an RF receiver comprising:
at least one RF receiving element comprising at least one antenna element receiving the reflected signal;
at least one mixer driven synchronously by the local RF source to generate one or more low-frequency output signals;
an analog-to-digital converter configured to digitize the low-frequency output signal and produce a digital data output at given time intervals;
a processing unit configured to:
execute a decision-making algorithm comprising a neural network model that receives the digital data output, wherein the neural network model is pre-trained and is subsequently fine-tuned or updated through transfer learning to accommodate a specific user profile to generate one or more output values; and
an alert system configured to activate upon detection of an object based on neural network output values satisfying one or more detection criteria.
2 . The wearable system of claim 1 , wherein the neural network model is trained on one or more of the low-frequency output signals to generate alerts.
3 . The wearable system of claim 2 , wherein the neural network model is a convolutional neural network (CNN) or recurrent neural network (RNN).
4 . The wearable system of claim 1 , wherein the local RF source generates an in-phase signal and a quadrature-phase signal.
5 . The wearable system of claim 1 , wherein the frequency of the local RF source is modulated over time.
6 . The wearable system of claim 1 , further comprising an inertial measurement unit configured to periodically measure inertial movements or vibration parameters of the wearable system and provide the periodically measured data to the decision-making algorithm.
7 . The wearable system of claim 6 , wherein the neural network model is trained on measured motion-related data, digital data output, or both.
8 . The wearable system of claim 1 , wherein the neural network model is configured to operate with an adaptive learning algorithm that dynamically updates during system operation based on real-time feedback and newly encountered scenarios.
9 . The wearable system of claim 1 , wherein the neural network model is executed entirely on the wearable system.
10 . The wearable system of claim 1 , wherein the processing unit progressively computes values of different layers of the neural network model during idle time.
11 . A wearable system adapted to detect an approaching object, comprising:
a radio frequency (RF) unit comprising:
a local RF source generating an output;
an RF transmitter comprising at least one antenna element driven in synchronization with the local RF source;
an RF receiver comprising:
at least one RF receiving element comprising at least one antenna element receiving a reflected signal;
at least one mixer driven synchronously by the local RF source to generate in-phase and quadrature-phase low-frequency output signals; and
an analog-to-digital converter configured to digitize the in-phase and quadrature-phase low-frequency output signals and produce digital signal data at given time intervals;
a processing unit configured to:
compute a spectral transform of the digital signal data to generate frequency-domain components; and
execute, on the wearable system, a neural network model trained on the frequency-domain components to generate one or more output values indicative of an approaching object and is subsequently fine-tuned or updated through transfer learning to accommodate a specific user profile; and
an alert system configured to activate based on the one or more output values satisfying one or more detection criteria.
12 . The wearable system of claim 11 , wherein the spectral transform comprises a discrete Fourier transform of the digital signal data to generate frequency bins having real and imaginary components, and wherein the neural network model is trained on, and configured to receive as input, the real and imaginary components of the frequency bins.
13 . The wearable system of claim 11 , wherein the processing unit progressively computes values of different layers of the neural network model during idle time.
14 . The wearable system of claim 11 , further comprising an inertial measurement unit configured to periodically measure inertial movements or vibration parameters of the wearable system and provide periodically measured data to the processing unit.
15 . The wearable system of claim 14 , wherein the neural network model is trained on measured motion-related data, the digital signal data, or both.
16 . The wearable system of claim 11 , wherein the processing unit progressively computes the spectral transform during idle time.
17 . The wearable system of claim 11 , wherein the frequency of the local RF source is modulated over time.
18 . A wearable system adapted to detect an approaching object, comprising:
a radio frequency (RF) unit comprising:
a local RF source generating an output;
an RF transmitter comprising at least one antenna element driven in synchronization with the local RF source;
an RF receiver comprising:
at least one RF receiving element comprising at least one antenna element receiving the reflected signal;
at least one mixer driven synchronously by the local RF source to generate one or more low-frequency output signals;
an analog-to-digital converter configured to digitize the low-frequency output signal and produce a digital data output at given time intervals;
a processing unit configured to:
execute a decision-making algorithm comprising a neural network model that receives the digital data output, wherein the neural network model is configured to Progressively compute values of different layers of the neural network model during idle time between successive time intervals of the analog-to-digital converter to generate one or more output values; and
an alert system configured to activate upon detection of an object based on neural network output values satisfying one or more detection criteria.