To address this challenge, we propose AutoDietary,
a wearable system to recognize food types by monitoring
eating process. The system is mainly composed of two parts:
(i) Embedded Hardware System and (ii) Smartphone
Application. An embedded hardware is developed to collect
and pre-process food intake data. The highlight is the necklacelike
acoustic sensors to pick up high-quality sound signals
of eating behaviors in a convenient and non-invasive manner.
The data are then transmitted via bluetooth to a smartphone,
where food types are recognized. We also developed an
application which not only aggregates food recognition
results but also provides the information in a user-friendly
way and offers suggestions on healthier eating, such as the
user should chew slower or should intake adequate hydration.
Specifically, food types can be distinguished from chewing
information, since the energy exerted in mastication basically
depends on the structural and the textural properties of the food
material and can be extracted from the chewing sound. Food
type recognition consists of several steps. The acoustic signals
are firstly framed. Then, the sound frames are processed
by the hidden Markov model (HMM) [4] based on the
Mel Frequency Cepstrum Coefficients [5] to detect the
chewing events. Moreover, we also detect fluid intake by
swallowing events. Then each event is processed to extract
several key features containing both time/frequency-domain
and non-linear information. A light-weight decision tree based
algorithm is adopted to recognize the type of food intake.
To evaluate our system, experiments are conducted involving
12 subjects to eat 7 different types of food. More than
4,000 food intake events are collected and evaluated. Results
show that the accuracy of identifying chewing/swallowing
events is 86.6%, based on which an accuracy of 84.9% can be
achieved for food type recognition. To classify liquid and solid
food, the accuracy can be up to 97.6% and 99.7%, respectively.
We also conducted a survey to investigate user experience of
AutoDietary. Results show that the current design (regarding
wear comfort and functionalities) is acceptable to most users.
By continuously monitoring eating behavior, AutoDietary can
provide customized suggestions to healthier eating habits, and