Brojeshwar Bhowmick

A multi-stage neural network aided system for detection of microcalcifications in digitized mammograms

We propose a multi-stage detection system for microcalcification. A connectionist online feature selection technique is used to identify a set of good features from a set of 87 features computed at a few randomly selected positive (calcified) and negative (normal) pixels. A neural network is then trained with the selected features. The network output is cleaned using connected component analysis and an algorithm for removing thin elongated structures. A measure of local density (called mountain potential) of the calcified points is then computed at every suspected pixel of these cleaned images and the peak of the mountain potential is used to classify mammograms as calcified or normal. The system is tested on a set of 17 mammograms comprising 10 abnormal and seven normal images which are not used in training and the system is found to perform very well. Moreover for each abnormal image, the system is able to locate the calcified regions quite accurately.

Detection and classification of eye state in IR camera for driver drowsiness identification

An eye detection and eye state (open/close) classification methodology for driver drowsiness idensification using IR camera has been presented in this paper. In this proposed methodology, otsu thresholding is used to extract face region. Eye localization is done by locating facial landmarks such as eyebrow and possible face center. Morphological operation and K-means is used for accurate eye segmentation. A hierarchial noise removal procedure is applied on the segmented image to get proper eye shape. Then a set of shape features are calculated and trained using nonlinear SVM to get the status of the eye. Experiment shows that the proposed methodology gives excellent segmentation results for both open eyes (both bright and dark pupil) and closed eyes and also classifies correctly.

Stereo Vision Based Pedestrians Detection and Distance Measurement for Automotive Application

This paper presents a compact and robust solution to use stereo vision to detect pedestrians as well as measure their instantaneous distances from the vehicle. This simplifies sensor requirements as the same video information (captured from two cameras) is used for detection of pedestrians as well as measurement of their corresponding distances. To apply stereo vision technique a novel approach has been made to locate identical points on human so that triangulation can be used for measuring distances. In addition to this, the paper also presents a novel technique used to make the pedestrian detection algorithm (based on well know Viola-Jones methodology) more robust by using tracking and assistive decision making (i.e. one camera helps other to eliminate noise). The proposed system is tested on standard x86 machine and gives good real time performance.

Mobiscan3D: A Low Cost Framework for Real Time Dense 3D Reconstruction on Mobile Devices

In this paper we propose a computationally inexpensive framework for dense 3D reconstruction on smart device platforms leveraging the feed from motion sensors. In contrast to other methods, we solely rely on the motion sensors to compute pair wise epipolar relationships. In particular, camera positions are obtained only through noisy mobile sensor data which is further optimized globally using iterative reweighted least squares. Rotations are also obtained using mobile sensors. Our method of obtaining camera motion reduce the processing time of the entire pipeline. We further use pair wise epipolar relationships along with normalized cross correlation and gradient information in a pair of images to obtain dense correspondences. The calibrations and correspondences are accurate enough for triangulation which in turn serve as a good initializer for final global bundle adjustment in near real time. Experimental results show that our method works reliably well for both indoor and outdoor scenes of different size and shapes without even utilizing mobile GPU.

An Application for Driver Drowsiness Identification based on Pupil Detection using IR Camera

A Driver drowsiness identification system has been proposed that generates alarms when driver falls asleep during driving. A number of different physical phenomena can be monitored and measured in order to detect drowsiness of driver in a vehicle. This paper presents a methodology for driver drowsiness identification using IR camera by detecting and tracking pupils. The face region is first determined first using euler number and template matching. Pupils are then located in the face region. In subsequent frames of video, pupils are tracked in order to find whether the eyes are open or closed. If eyes are closed for several consecutive frames then it is concluded that the driver is fatigued and alarm is generated.

Quantification of balance in single limb stance using kinect

This paper presents a novel single limb body balance analysis system which will aid medical practitioners to analyze crucial factor for fall risk minimization, injury prevention, fitness and rehabilitation programs. We use skeleton data obtained from Microsoft Kinect which captures full human body as well as ensures users privacy. A new eigen vector based curvature analysis algorithm is developed to compute single limb stance (SLS) duration on the skeleton data. Two parameters vibration-jitter and force per unit mass (FPUM) are derived for each body part to assess postural stability during SLS. Experimental results show the efficacy of our system to apply it in medical domain.

Accurate upper body rehabilitation system using kinect

The growing importance of Kinect as a tool for clinical assessment and rehabilitation is due to its portability, low cost and markerless system for human motion capture. However, the accuracy of Kinect in measuring three-dimensional body joint center locations often fails to meet clinical standards of accuracy when compared to marker-based motion capture systems such as Vicon. The length of the body segment connecting any two joints, measured as the distance between three-dimensional Kinect skeleton joint coordinates, has been observed to vary with time. The orientation of the line connecting adjoining Kinect skeletal coordinates has also been seen to differ from the actual orientation of the physical body segment. Hence we have proposed an optimization method that utilizes Kinect Depth and RGB information to search for the joint center location that satisfies constraints on body segment length and as well as orientation. An experimental study have been carried out on ten healthy participants performing upper body range of motion exercises. The results report 72% reduction in body segment length variance and 2° improvement in Range of Motion (ROM) angle hence enabling to more accurate measurements for upper limb exercises.The growing importance of Kinect as a tool for clinical assessment and rehabilitation is due to its portability, low cost and markerless system for human motion capture. However, the accuracy of Kinect in measuring three-dimensional body joint center locations often fails to meet clinical standards of accuracy when compared to marker-based motion capture systems such as Vicon. The length of the body segment connecting any two joints, measured as the distance between three-dimensional Kinect skeleton joint coordinates, has been observed to vary with time. The orientation of the line connecting adjoining Kinect skeletal coordinates has also been seen to differ from the actual orientation of the physical body segment. Hence we have proposed an optimization method that utilizes Kinect Depth and RGB information to search for the joint center location that satisfies constraints on body segment length and as well as orientation. An experimental study have been carried out on ten healthy participants performing upper body range of motion exercises. The results report 72% reduction in body segment length variance and 2° improvement in Range of Motion (ROM) angle hence enabling to more accurate measurements for upper limb exercises.

A Kalman Filter Based Approach to De-noise the Stereo Vision Based Pedestrian Position Estimation

This paper presents a methodology of using Kalman filter to minimize the error in stereo vision based distance measurement data (3D position of pedestrians). In stereo vision, little point mis-correspondence leads to a very bad estimate of depth during triangulation. There are robust correspondence algorithms but all of them suffer from algorithm complexity affecting the time performance. If simple correspondence algorithms are used that gave good real time performance, then the results suffer from erroneous depth measurement. In this paper, we have applied a predictive-corrective model using Kalman filter on the erroneous depth measurement. Being applied in time domain as compared to stereo image domain, the proposed approach has much less algorithm complexity and hence gives good real-time performance. The results also show that the proposed algorithm is able to significantly reduce the measurement noise without affecting the pedestrian tracking ability.

An application for retrieval of frames from a laparoscopic surgical video based on image of query instrument

This paper describes a practical and reliable solution/approach to achieve automated retrieval of surgical instruments used in laparoscopic surgery. The central goal is to locate particular video frames containing intended information which can be used for analysis and diagnosis. In this paper, a practical system is proposed where the users need not manually search the candidate frames in the entire video. Instead, users can give any query object (in image format) and the frames containing the object will be retrieved. Given an object image, the method extracts features like color and shape of objects in each frame of the laparoscopic video and compare with the input image feature to retrieve the frames containing the desired instrument. The system can recognize the instrument in 91% cases but does not give any false alarm. Experimental results are presented to show the feasibility of the proposed application.

A System for Near Real-Time 3D Reconstruction from Multi-view Using 4G Enabled Mobile

Rapid or real time dense 3D reconstruction of indoor or outdoor model using mobile phone camera has become an active research topic because of the importance of 3D models in different application like remote monitoring system and visualization. In this paper we propose an application as well a system architecture which captures images using a hand held mobile Smart phone and generate an accurate and dense 3D model in a connected server using parallel computation. The method performs near real time accurate dense 3D model generation from streaming data. The application is capable of adaptive data transmission based on available network bandwidth. A comparison is presented for different amount of data transmission. The main computation is performed on a high end server system that reconstructs an accurate and complete 3D model with registration between all created models.