Skip to Main Content

Strategic Business Insights (SBI) logo

Novel Identification Tech Featured Pattern: P1447 January 2020

Author: Guy Garrud (Send us feedback.)

Advances in neural networks and sensors can enable a variety of devices to identify objects and activities.

Abstracts in this Pattern:

Most existing object-recognition systems rely on visual data and machine learning to identify what a camera is looking at. However, visual data have a variety of drawbacks, including that they lack information about texture, softness, and weight distribution—information that is vital for enabling robots to handle everyday objects effectively. Researchers at the Korea Advanced Institute of Science and Technology (KAIST; Daejeon, South Korea) have developed machine-learning software that can identify an object without the use of visual data or special hardware. The Knocker ( software requires users to tap an object with a smartphone and then uses data from the smartphone's sensors—including accelerometers, gyroscopes, and microphones—to classify and identify the object.

Wearable technology can provide substantial information about objects and activities. For example, researchers at the Carnegie Mellon University (CMU; Pittsburgh, Pennsylvania) Human-Computer Interaction Institute are using data from the sensors in smartwatches to identify what wearers are doing with their hands. The researchers made a few alterations to a smartwatch's operating system, enabling the use of accelerometer data and even bioacoustic sounds to distinguish among 25 common activities, including petting an animal, pouring a drink, typing, and washing dishes. The researchers report that their method is 95% accurate if the smartwatch is on the wearer's dominant arm. And researchers at the Massachusetts Institute of Technology (MIT; Cambridge, Massachusetts) have developed a sensor-enabled glove that provides object-related information. The glove uses a force-sensitive film on its palm and fingers and a network of conductive silver threads to "detect the weight and shape of an object its wearer is holding, as well as the pressure created as the hand moves." Data from the sensors pass to a neural network that finds patterns to identify, for example, whether an object has edges or is spherical.

Further advances in object recognition are important for numerous applications, including dexterous robots and a variety of health-care and assistive technologies. Such advances will also benefit wearable devices by enabling them to determine with a high degree of accuracy what wearers are doing.