WEARABLE COMPUTING

Designing and Sharing Activity-Recognition Systems Across Platforms

BY, KRISHNA PRASAD C E7 A ROLL NO: 33 GUIDE:RASHIDA

OVERVIEW
INTRODUCTION WEARABLE COMPUTERS WERABLE COMPUTING WEARABLE ACTIVITY RECOGNITION ACTIVITY RECOGNITION CHAIN SHARING ACTIVITY-RECOGNITIONSYSTEM CONCLUSION AND OVERLOOK REFERENCE

INTRODUCTION
In robotics, activity-recognition systems can be used to label large robot-generated activity data sets.

It enables activity-aware humanrobot interactions (HRIs). It also opens ways to self-learning autonomous robots. The recognition of human activities from body-worn sensors is a key paradigm in wearable computing. In this field, the variability in human activities, sensor deployment characteristics, and application domains has led to the development of best practices and methods to enhance the robustness of activity-recognition systems. Finally, the current challenges in wearable activity recognition are outlined.

WEARABLE COMPUTER
Wearable computers are computers that are worn on the body. They have been applied to areas such as behavioral modeling, health monitoring systems, information technologies and media development. Wearable computers are especially useful for applications that require computational support while the user's hands, voice, eyes or attention are actively engaged with the physical environment. Areas of study include user interface design, augmented reality, pattern recognition, use of wearables for specific applications or disabilities, electronic textiles and fashion design.

WEARABLE COMPUTER
A wearable computer is a computer that is subsumed into the personal space of the user, controlled by the user, and has both operational and interactional constancy. It is a device that is always with the user, and into which the user can always enter commands and execute a set of such entered commands, The wearable computer is more than just a wristwatch or regular eyeglasses: it has the full functionality of a computer system but in addition to being a fully featured computer.

Overall System Block Diagram : Overall System Block Diagram :

Input Device Display Device

Com port

VGA out Network card

Wireless Wireless Network Network

Frame grabber

Video Camera

Main Unit

Parallel port

Back plane

Low Power Indicator

Power Supply

Land Warrior Breakdown

A. Eye piece display o Friendly Positions o Maps o Mission data B&C. Mic and Earphones D.Input device that acts as a mouse. E.Laser range finder to send data about targets to others F.Thermal Sight for night vision G. Video Camera o Send pics and clips o Shoot around corners

INPUT DEVICE

You hold up your left hand, fingers pointing to the right. The system recognizes that you want to make a call, and projects a dialing pad onto your fingers. You tap the virtual keypad with your right hand to dial the call.

WEARABLE COMPUTING
Wearable computing, as originally presented by Mann in 1996, emphasized a shift in computing paradigm [22].Computers would no longer be machines separate from the persons using them. Instead, they would become an unobtrusive extension of our very bodies, providing us with additional ubiquitous sensing, feedback, and computational capabilities. Wearable computing never considered implanting sensors or chips into the body. Rather, it emphasizes the view that clothing,which has become an extension of our natural skin, would be the substrate that technology could disappear into. The prevalence of mobile phones now offers an additional vector for on-body sensing and computing [23].Mann [24] and Starner et al. [25] were among the first to show that complex contextual information can be obtained by interpreting on-body sensor data and that this would lead to novel adaptive applications. A wearable system can perceive activities, defined here to include both gestures and behaviors, from a first-person perspective. This leads to new forms of applications known as activity based computing or interaction-based computing [20],[21]. Such applications can offer information or assistance pro-actively based on the users situation as well as support explicit interaction in unobtrusive ways through natural gestures or body movements.

Wearable Activity Recognition

Activity and gesture recognition are generally tackled as a problem of learning by demonstration . The user is instrumented with the selected sensors and is put into a situation where he performs the activities and gestures of interest. The sensor data are acquired with ground-truth annotations describing what the user performs or experiences. The resulting data set is used to train the recognition system and test its performance. The training process consists of identifying the mapping between the users activities or gestures and the corresponding sensor signals. Some terminology commonly used in wearable activity recognition differs from the one used in robotics. Annotation or Labeling: This is the process by which the experimenter manually provides ground-truth information about the activities of the subject, generally, when collecting an activity data set. Recognition or Spotting: This is the actual machine identification of an activity in the data sensor stream. Activities are said to be recognized or spotted.

Sensors for Activity Recognition

Sensors are used to acquire signals related to the usersactivities or gestures. User comfort is paramount. Thus,the sensors must be small, unobtrusive, and ideally invisible to the outside. The sensors are selected according to a tradeoff between wearability, computational needs, power usage, communication requirements, and information content for the activities and contexts of interest.Common sensor modalities are body-worn accelerometers and IMUs. Accelerometers are extremely small and have low power. The IMUs contain accelerometers, magnetometers, and gyroscopes, which allow to sense the orientation of the device with respect to a reference. The IMUs are typically placed on each body segment and allow to reconstruct a body model of the user. On-body microphones are also successfully used for activity recognition, as many human activities generate characteristic sound.Today, the trend goes toward an increased use of multiple multimodal sensors, as this tends to increase recognition performance (see [30]). Wearable systems are also complemented by object integrated and ambient sensors.

Sharing Activity-Recognition Systems

Human activity recognition in wearable computing is challenging because of a large variability in the mapping of sensor signals to activity classes. This variability has multiple origins, which is shown below. Semantically identical action primitives can be executed in a large number of ways . This is referred to as intra user variability. These variations arise from personal preferences. Although different persons may be considered as robots of identical make, in practice, there is an even higher variability in action-motor strategies between users (inter user variability) than for a single user. The placement of the sensors on body cannot be done with a high precision, especially when the users deploy the sensors themselves.

Figure 6 illustrates the effect of sensor placement on the projection of sensor data into the feature space.l Abstracting the specific environment in which the system can recognize activities is important to ensure costeffective deployment on a large scale. Thus, activity-recognition methods should work for a generic class of problems (e.g., in any smart home) rather than a specific instance of the problem class (e.g., a specific smart home). To further increase unobtrusiveness, we argue in the project OPPORTUNITY (http://www.opportunity-project.org, a Future and Emerging Technologies project funded by the European Commission in the seventh framework programme for research) to use opportunistically discovered sensors for activity recognition [17], [50]. Thus, the available sensor configuration depends on the sensorized objects users take with themselves, on the smart clothing they wear, and on the environment in which they are located. For each sensor kind and placement, there is a different sensor-signal to activityclass mapping that an opportunistic activity-recognition system should be able to abstract.The wearable computing community has developed best practices and novel methods to deal with some forms of variability. In the following subsections, we present a selection of methods developed by various groups and ours. To share an ARC, there must be a common representation at some stage in the recognition chain.We organize the methods along the level at which the methods assume a common representation. We describe methods operating at the sensor, feature, classifier, and reasoning levels .

SENSOR LEVEL SHARING

This level focuses on training an ARC on the first platform and reusing it on the second platform. This assumes that the sensor-signal to activity-class mappings are statistically identical on two platforms. Training an ARC on one system is referred to as a user-specific system, and it is known to show degraded performance when deployed to another. The best practice to realize an ARC that generalizes to new situations consists in training it on a data set containing the variability to be seen when the system is deployed. By collecting a data set from multiple users, the ARC can be trained to be user independent. By collecting a data set comprising multiple on-body sensor positions, the ARC can be trained to be independent of sensor placement.

SENSOR LEVEL SHARING(cont) The previous approach requires to foresee all the variations likely to be encountered at run time. Thus, we proposed an unsupervised self-calibration approach that removes this requirement. The self-calibration approach operates as follows: the ARC continuously operates and recognizes the occurrence of activities/gestures. upon detection of an activity/gesture, the corresponding sensor data is stored as training data. the classifiers are retrained, including this new training data, using an incremental learning algorithm. Thus, the activity models are optimized upon each activity instance to better model that activity.

FEATURE LEVEL SHARING

At this level, the ARC devised for the first platform is translated to the second platform from the feature level onwards. The use case for sharing ARCs at this level include systems where the sensor modalities on the two platforms do not coincide or show large on-body displacement for which a placement-independent ARC cannot be envisioned. They show that features that are robust to on-body displacement can be designed using body models and by fusing multiple sensors such as an accelerometer and a gyroscope. They also show that a specific sensor modality (magnetic field sensor) can be replaced by another specific modality (gyroscope). use on body sensor placement self-characterization as a way to select, among a number of preprogrammed ARCs, the one most suited for the detected sensor placement. in robotics, data from different sensors can be converted into identical abstract representations.

CLASSIFIER LEVEL SHARING

Transfer learning allows to translate a classification problem from one feature space to another and was used to transfer perceptual categories across modalities in biological and artificial systems . Conceptually, transfer learning may thus be used to translate the capability to recognize activities from one platform to another without enforcing a similar input space (i.e., sensors and features). The transfer does not affect higher-level reasoning. Practical principles allowing a system A to confer activity recognition capabilities to another system B are outlined in . A new system is deployed in the users personal area network comprising a set of unknown new sensors SB (on body and/or in the users surroundings) and an untrained ARCB. As the user performs activities, the ARCA recognizes them and broadcasts this information.

SYMBOLIC LEVEL SHARING

The reasoning program to infer higher level activities from spotted action primitives is shared between platforms. As the environment in which the two platforms operate may lead to the detection of semantically different action primitives, a direct transfer of the reasoning is not always possible. Carrying out a prior concept matching can address this.

The system first automatically finds how sensor activations in different environments relate to identical higherlevel concepts using statistical approaches. A recognition system can also learn internal hierarchical representation of activities or concepts [60], upon which reasoning is performed. Hu et al.

further report on using Web mining to match concepts [61]. Advances in merging concepts in ontologies , [support the transfer of activity-recognition reasoning across different conceptual spaces. In robotics, these principles may allow the robots to exchange the knowledge they have individually gained about the world. This may be especially relevant when principles of autonomous mental development are used, as robots can develop distinct world representations according to their capabilities.