Apparatus and method for performing collision analysis
An electronic device may be provided to perform collision analysis. The electronic device may be configured to: receive sensor data from an inertial-measuring-unit (IMU) sensor; detect that a collision event has occurred based on the sensor data and detect collision data from the sensor data; input the collision data into a collision classification model that is trained to predict a collision material type and a collision intensity of the collision event; and determine an action to be performed by the electronic device based on the collision material type and the collision intensity of the collision event.
1 . A hand-held electronic device for performing collision analysis, the hand-held electronic device comprising:
an inertial-measuring-unit (IMU) sensor configured to collect sensor data about movements of the hand-held electronic device, which are caused by an impact applied directly on an enclosure of the hand-held electronic device;
at least memory storing instructions; and
at least one processor configured to execute the instructions to:
receive the sensor data from the IMU sensor;
detect that a collision event has occurred on the hand-held electronic device, based on the sensor data and detect collision data from the sensor data;
input the collision data into a collision classification model that is trained to predict a collision material type and a collision intensity of the collision event, wherein the collision material type corresponds to a floor or an object on which the hand-held electronic device is dropped; and
determine a service action to be performed by the hand-held electronic device based on the collision material type and the collision intensity of the collision event,
wherein the collision classification model comprises a first neural network that constitutes a collision material classification model and a second neural network that constitutes a collision intensity classification model, and
wherein the collision material classification model and the collision intensity classification model share an input layer and hidden layers, and have separate output layers configured to output information of the collision material type and the collision intensity, respectively.
2 . The hand-held electronic device of claim 1 , wherein the collision classification model is trained for hierarchical multi-label classification.
3 . The hand-held electronic device of claim 1 , wherein the collision classification model is trained based on a combination of a material classification task loss between the predicted collision material type and a ground-truth material type label for a material classification task, an intensity classification task loss between the predicted collision intensity and a ground-truth intensity label for an intensity classification task, and a task consistency loss that represents a degree of match between the predicted collision material type and the predicted collision intensity.
4 . The hand-held electronic device of claim 3 , wherein the task consistency loss is computed by comparing a pair of the predicted collision material type and the predicted collision intensity with ground-truth inter-class relationships between sample collision material types and sample collision intensities.
5 . The hand-held electronic device of claim 1 , wherein the collision classification model is trained by:
accessing the sensor data from previous collision events with a plurality of material types;
generating a plurality of training examples by extracting portions of the sensor data for each of the previous collision events, wherein the extracted portions have a same duration but different start times and end times, and wherein each of the plurality of training examples is labeled with a material type and a collision intensity type of the previous collision events; and
training the collision classification model with the plurality of training examples.
6 . The hand-held electronic device of claim 1 , wherein the service action comprises at least one of calling an emergency contact or notifying an administrator or a user of the hand-held electronic device.
7 . The hand-held electronic device of claim 1 , further comprising a user interface configured to receive user feedback about the collision event,
wherein the user feedback comprises user inputs of the collision material type and the collision intensity of the collision event, and
wherein the collision classification model is updated based on the user feedback.
8 . A method for performing collision analysis by a hand-held electronic device, the method comprising:
receiving sensor data from an inertial-measuring-unit (IMU) sensor, the sensor data being related to movements of the hand-held electronic device, which are caused by an impact applied directly on an enclosure of the hand-held electronic device;
detecting that a collision event has occurred on the hand-held electronic device, based on the sensor data and detecting collision data from the sensor data;
inputting the collision data into a collision classification model that is trained to predict a collision material type and a collision intensity of the collision event, wherein the collision material type corresponds to a floor or an object on which the hand-held electronic device is dropped; and
determining a service action to be performed by the hand-held electronic device based on the collision material type and the collision intensity of the collision event,
wherein the collision classification model comprises a first neural network that constitutes a collision material classification model and a second neural network that constitutes a collision intensity classification model, and
wherein the collision material classification model and the collision intensity classification model share an input layer and hidden layers, and have separate output layers configured to output information of the collision material type and the collision intensity, respectively.
9 . The method of claim 8 , wherein the collision classification model is trained for hierarchical multi-label classification.
10 . The method of claim 8 , wherein the collision classification model is trained based on a combination of a material classification task loss between the predicted collision material type and a ground-truth material type label for a material classification task, an intensity classification task loss between the predicted collision intensity and a ground-truth intensity label for an intensity classification task, and a task consistency loss that represents a degree of match between the predicted collision material type and the predicted collision intensity.
11 . The method of claim 10 , wherein the task consistency loss is computed by comparing a pair of the predicted collision material type and the predicted collision intensity with ground-truth inter-class relationships between sample collision material types and sample collision intensities.
12 . The method of claim 8 , wherein the collision classification model is trained by:
accessing the sensor data from previous collision events with a plurality of material types;
generating a plurality of training examples by extracting portions of the sensor data for each of the previous collision events, wherein the extracted portions have a same duration but different start times and end times, and wherein each of the plurality of training examples is labeled with a material type and a collision intensity type of the previous collision events; and
training the collision classification model with the plurality of training examples.
13 . The method of claim 8 , wherein the service action comprises at least one of calling an emergency contact, notifying an administrator or a user of the hand-held electronic device, and logging the collision event.
14 . The method of claim 8 , further comprising:
receiving user feedback about the collision event, via a user interface, wherein the user feedback comprises user inputs of the collision material type and the collision intensity of the collision event, and
updating the collision classification model based on the user feedback.
15 . A non-transitory computer readable storage medium that stores instructions to be executed by at least one processor included in a hand-held electronic device to perform a method for performing collision analysis about the hand-held electronic device, the method comprising:
receiving sensor data from an inertial-measuring-unit (IMU) sensor, the sensor data being related to movements of the hand-held electronic device, which are caused by an impact applied directly on an enclosure of the hand-held electronic device;
detecting that a collision event has occurred on the hand-held electronic device, based on the sensor data and detecting collision data from the sensor data;
inputting the collision data into a collision classification model that is trained to predict a collision material type and a collision intensity of the collision event, wherein the collision material type corresponds to a floor or an object on which the hand-held electronic device is dropped; and
determining a service action to be performed by the hand-held electronic device based on the collision material type and the collision intensity of the collision event,
wherein the collision classification model comprises a first neural network that constitutes a collision material classification model and a second neural network that constitutes a collision intensity classification model, and
wherein the collision material classification model and the collision intensity classification model share an input layer and hidden layers, and have separate output layers configured to output information of the collision material type and the collision intensity, respectively.