As stated by the World Health Organization (WHO) “Nutrition is coming to the fore as a major modifiable determinant of chronic disease, with scientific evidence increasingly supporting the view that alterations in diet have strong effects, both positive and negative, on health throughout life” [1]. It is therefore of key importance to find efficient and solid methodologies to study eating behavior and food intake in order to help reduce potential long-term health problems caused by unhealthy diets. Past research on eating behaviors and attitudes relies intensively on self-reporting tools, such as 24-h recalls, food records (food diaries) and food frequency questionnaires (FFQ; [2,3,4]). However, there is an increasing understanding of the limitations of this classical approach to studying eating behaviors and attitudes. One of the major limitations of this approach is that self-reporting tools rely on participants’ recall, which may be inaccurate or biased (especially when studying the actual amount of food or liquid intake [5]). Recall biases can be caused by demand characteristics, which are cues that may indicate the study aims to participants, leading them to change their behaviors or responses based on what they think the research is about [6], or more generally by the desire to comply with social norms and expectations when it comes to food intake [7,8].

There is growing interest in identifying technologies able to improve the quality and validity of data collected to advance nutrition science. Such technologies should enable eating behavior to be measured passively (i.e., without requiring action or mental effort on the part of the users), objectively and reliably in realistic contexts. To maximize the efficiency of real-life measurement, it is vital to develop technologies that capture eating behavior patterns in a low-cost, unobtrusive and easy-to-analyze way. For real-world practicality, the technologies should be comfortable and acceptable so that they can be used in naturalistic settings for extended periods while respecting the users’ privacy.

For “eating/drinking activity detection”, many systems have been reported that measure eating- and drinking-related motions. In particular, many papers reported measuring these actions using motion sensors such as inertial sensor modules (i.e., inertial measurement units or IMUs). IMUs typically consist of various sensors such as an accelerator, gyroscope and magnetometer. These sensors are embedded in smartphones and wearable devices such as smartwatches. In [16], researchers collected IMU signals with off-the-shelf smartwatches to identify hand-based eating and drinking-related activities. In this case, participants wore smartwatches on their preferred wrists. Other studies have employed IMUs worn on the wrist, upper arm, head, neck and combinations thereof [17,18,19,20]. IMUs worn on the wrist or upper arms can collect movement data relatively unobtrusively during natural eating activities such as lifting food or bringing utensils to the mouth. Recent research has also improved IMUs that are attached to the head or neck, combining sensors with glasses or necklaces so that they are less bulky and users are not aware that they are being worn. Besides IMUs, proximity sensors, piezoelectric sensors and radar sensors are also used to detect hand-to-mouth gestures or jawbone movements [21,22,23]. Pressure sensors are used to measure eating activity as well. For instance, in [24], eating activities and the amount of consumed food are measured by a pressure-sensitive tablecloth and tray. These devices provide information on food-intake-related actions such as cutting, scooping, stirring or the identification of the plate or container on which the action is executed and allow the tracking of weight changes of plates and containers. Microphones, RGB-D images and video cameras are also used to detect eating and drinking-related motions.

2.2. Bite, Chewing or Swallowing Detection