Please note! This essay has been submitted by a student.
Nose tracking: – The nose feature is defined as the point on the nose surface that is the closest to the camera. This point is termed the tip of the nose. Due to the symmetry and the convex shape of the nose, the nose feature is always visible in the camera, and it stays almost the same during the rotations of the head. It also does not change much with head moving towards and from the camera. Thus, the nose tip defined above can always be located. This is very an important property of the nose which does not hold for any other facial feature.
Weiwei Zhang etl. proposed a yawning detection system that consists of a face detector, a nose detector, a nose tracker and a yawning detector. Deep learning algorithms are refining for detecting driver face area and nose location. A nose tracking algorithm that combines a Kalman filter with a dedicated open-source TLD (Track-Learning-Detection) tracker is developed to achieve robust tracking results under dynamic driving conditions.
The computational framework is developed for robust driver yawning detection under various
illumination conditions of daytime driving. The system contains three major components, a face detector, a nose detector, a nose tracker and a yawning detector. The face detector uses a deep convolutional neural network (CNN) to locate the face area in a driver image, and another CNN is developed to detect driver’s nose within the detected face area. the multi-stage yawning detection system, each stage’s performance is depended on results generated by the previous stages. The results generated by the nose detector are used as input to the TLD-based nose tracker to estimate the motion in the following frames, track the change of dynamic appearance of driver’s face expression in video imagery, identify nose bounding box, and generate nose tracking confidence in each frame. The nose position in the previous frame generated by Kalman filtering to achieve efficient and effective results. For yawning detection, we use gradient features around the mouth corners, and the gradient value has significant changes during yawning. the nose position in the previous frame generated by Kalman filtering to achieve efficient and effective results. For yawning detection, we use gradient features around the mouth corners, and the gradient value has significant changes during yawning.
S. Waphare et al. proposed the implementation of 2 novel algorithms named Surge-spiralx and Surge-castx on sniffer robot for odor plume tracking(i.e. The e-nose takes in samples of the environment air, where the sensors translate the odor concentration into electrical signals which are then processed to identify the odor.) in a laminar wind environment.
The Sniffer Robot Developed is based on Atmega32 platform which acts as the microcontroller and has useful functional units such as inbuilt ADC, PWM generator etc. The Sniffer Robot is provided with an anemometer for detection of wind tunnel angle which consists of 4 NTC thermistors separated by a ‘+’ shaped diaphragm. It has 3 Metal Oxide Gas sensors of TGS series, which are spatially separated and mounted such that one faces the front side; one towards left and the remaining one face right. The Gas Sensors used here have been used as analog sensors, so that the sensor data of the 3 sensors can be compared and the decision can be made based on it which is necessary for Surge-Spiralx and Surge-Castx algorithms.
In Surge-Spiral algorithm, Gas sensors are used only to indicate whether the odor is present or absent. In Surge-Spiralx Gas sensors are used an analog manner and the 3 ADC values are compared in order to take a decision. The gas sensors are quite noisy, hence the sensitivity needs to be kept at minimum and techniques such as suction fan need to be implemented in order to improve performance. In Surge-Spiralx algorithm, the upwind surge from Surge- Spiral is replaced by the differential surge. When the robot is inside odor plume it moves upwind and at the same time, it runs the correction loop.
The algorithms are developed and they have shown very good performance in terms of success ratio, while Surge-Spiralx algorithm having less distance overhead.
C. Mouth Tracking: -The term mouth tracking is used to include lip tracking as; the lips are a component of the mouth which contains other vital cues describing the mouth (i.e. tongue, teeth, oral cavity). The lips however, act as an invaluable feature for tracking the mouth as in many cases the labial area gives a very good line of demarcation between the mouth and the face background.
Jie Cheng et al. proposed a novel approach for real-time mouth tracking and 3D reconstruction. This method comprises two successive processing stages. In the first stage, an AdaBoost learning algorithm and a Kalman filter are used to detect and track the mouth region in real-time under a complex background. In the second stage, the resultant 2D position of the mouth is used to determine the region where the 3D shape is reconstructed by use of a digital fringe projection and modified Fourier transform method.
AdaBoost algorithm is a kind of general-purpose learning algorithms used to enhance the performance of a weak classifier Error! Reference source not found. Another important characteristic of Viola and Jones’ object detection framework is a cascade of classifiers constructed to achieve improved detection performance while reducing computation time.
Kalman filter is one of the most widely used mathematical tools for stochastic estimation from noisy sensor measurements. It uses a predictor-corrector type estimator that minimizes the estimated error covariance. The Kalman filter is used to smooth the temporal trajectories of the center position and size of the rectangle. It can also be used to predict the change of these quantities. The mouth region is detected, the 3D shape of the mouth in the detected region is dynamically reconstructed. Digital fringe projection is a fast full-field 3D shape measurement technique. With a digital-light-processing (DLP) projector, different fringe patterns can be programmed into the color channels of the projector, thus achieving a high fringe projection speed.
The mouth is tracked in 2D using our proposed hierarchical 2D mouth positioning method, which is a combination of Viola and Jones’ rapid object detection framework and Kalman filter. The 3D shape is reconstructed using the high-speed 3D reconstruction system based on a digital fringe projection and a modified Fourier transform method. The main grant of this paper is the real-time dense 3D reconstruction of the mouth region, which can be useful in many applications, such as lip-reading, biometrics, 3D animation, etc.
Sunil S. Morade et al. proposed a novel active contour guided geometrical feature extraction approach for lip reading. Three active contour approaches are snake, the region scalable fitting energy method and localized active contour model.
The mouth is located, algorithms can be used for lip contour estimation. Here three methods of active contour for boundary extraction are compared. One of the most common methods of active contour models is snakes. The Second method is based on Region scalable fitting energy method (RCFE). The Third method is localized active contour model (LACM) which depends upon local( region)information. The snake is an energy minimizing spline guided by external forces and influenced by image. Snakes are active control models. The snake technique was first introduced by Kass, Witkin, and Terzoupoulos. Snake algorithm adapts the algorithm of Williams and Shah.
E = ∫▒〖(∝(s) E_cont+β(s) E_(curve )+γ(s) E_image )ds (1) 〗
Minimization of Region-Scalable Fitting Energy method was introduced by C. Li et al., used for biomedical image segmentation. The term is derived from the data fitting energy; this term plays a key role in the model, since it is responsible for driving the active contour toward object boundaries fitting term.
Localized contours are capable of segmenting objects with heterogeneous feature profiles that would be difficult to capture correctly using a standard global method. PCA is used to obtain optimal combinations from the statical point of view. In PCA technique the first variance of input matrix is calculated (for a combination of the geometric parameter). The Principal component analysis is applied for a combination of lip parameters. Eigen vectors associated with higher eigen values are selected.
These approaches are adopted for salient geometrical feature calculation. The experimentations were carried out for lip feature extraction which is useful for visual speech recognition. A joint feature model, obtained by combination inner area, height and width has been proposed.