Prof. K. T. Talele

A vision-based real-time driver fatigue detection system is proposed for increase in driver safety. Driver fatigue is evaluated on the basis of two parameters: rate of eye closure and yawning. Images of the driver are taken through a camera. Each individual image is checked for the face, eyes and the mouth. If they do not comply with the rules set for fatigue detection then, there is an immediate warning given in the form of an alarm.  The system is tested on a hardware platform consisting of a TI OMAP 3530 chip along with other supporting peripherals. The experimental results obtained so far show great promise in the near future. The system can reach 20 frames per second for eye tracking. Efficiency of the system is near 90% in standard light conditions. However, the system may give erroneous results when the illumination level goes low.

The objective of the project is to design a speech operated motion tracker for detection of motion. Our tracker is an essential requirement of any surveillance system. Nowadays CCTV cameras are used for security purposes. These cameras can even track the object that is in motion but as they are .xed at a particular place.In this work we present an operational computer vision system for real-time detection and tracking of object in motion. The system captures video of a scene and identi.es moving objects using wavelet segmentations, oper ated by Speaker through speech. We have implemented speech recognition using Wavelet Decomposed LPC which gives 80 percent accuracy. Using Wavelet segmentation we were able to detect motion even in varying brightness, this gives an advantage of tracking motion in the regions of low brightness. Motion tracker can be used in surveillance areas where the CCTV cameras fail, Motion tracker will be able to follow the object until a clear image of the object is retrieved.

We present a method to aggregate certain featuresin a particular class of objects into a visual dictionary. They are compared to the image given by the user to be recognized. The invariant features are extracted which become the reference for the input image detection. After this, the image descriptor is used for classification and then consequent recognition. Stable local feature detection and representation is a fundamental component of many image registration and object recognition algorithms. The SIFT algorithm as being the most resistant to common image deformations, a general trainable system for object detection in static images has been proposed that examines and improves upon the local image descriptor. The salient aspects of the image gradient are decoded in the feature point’s neighborhood. The features are invariant to image scale and rotation, and are shown to provide robust matching across a a substantial range of affine distortion, change in 3D viewpoint, addition of noise, and change in illumination. The recognition proceeds by matching individual features to a database of features from known objects using a fast nearest-neighbor algorithm. This approach to recognition can robustly identify objects among clutter and occlusion. We apply SIFT’s smoothed weighted histograms to make it robust to image deformations and rotations. Normalization and scaling is inherently done. This system can be applied to face, people, car, objects like cycles, bikes for road systems; and aeroplanes, ammunition, for security; retrieval for image mining of databases, etc. This has been implemented for the first time in Visual C++ using OpenCV.

This project involves the development of a bimodal emotion recognition system using facial expressions and body gestures. Affect sensing by machines has been argued as an essential part of next-generation human-computer interaction (HCI). To this end, in the recent years a large number of studies have been conducted, which report automatic recognition of emotion as a difficult, but feasible task. In this project, we have extended this idea using multimodal emotion detection technique to recognize extreme emotions so that this concept can be used in various real time applications successfully. This project is made in view of its use in rating advertisements by detecting and recognizing emotions of a person viewing the advertisement. This analysis can be then used to rate the advertisement and to make the required changes to increase its success rate.

26th July 2005 saw the heaviest rainfall ever in the city of Mumbai, as a result of which the city experienced the worst floods in history. An important factor that led to the loss was that the government agencies were unable to relay warnings at appropriate time due to lack of resources. The proposed system takes into account the rain-rate along with the topography of an area to predict possibility of floods. It is based on the fact that in case of rainfall the low-lying regions are more likely to get flooded, the water then eventually moving to the other areas. This can give a fair idea of the time in which the any region under consideration is likely to flood. This concept has been implemented by dividing the area into watersheds. With this technique, it is possible to monitor rain-rate in real time and reduce the prediction time.

Our application presents simple but yet efficient technique of human face sensing in distributed database. Our system is embedding face sensing techniques with distributed database management concept which will make criminal or missing person identification process very fast and time efficient. We have used a distributed database system which is a network of two or more SQL Databases that reside on one or more machines. Each database will have millions of face images. After getting an image at server, it will apply face sensing technique based on Template Matching algorithm. It detects human face in different scales, various poses, different expressions, lighting conditions, and orientation. We will use parallel processing in which application will call each database at time for matching images and will receive reply from individual databases. Application displays valuable information about the person whose image has been identified or recognized with quick service. Experimental results show the proposed system obtains competitive results and improves detection performance substantially.

Number Plate Recognition technology is gaining popularity in security and traffic installations. The technology concept assumes that all vehicles already have the identity displayed (the number plate) so no additional transmitter or responder is required to be installed on the car. The system uses the image of the front or rear of the vehicle, then the image-processing software analyses the images and extracts the plate information. The system takes an image of the vehicle as input, preprocesses this image using various methods to get the number plate region. Then characters of the plate are segmented and matched with the template to give number plate as the final output. This system’s significant advantage is that it does not need any installation per car. The application is developed using MS Visual Studio 2008 and OpenCV 2.1 and will effect in easier and accurate detection of the number plate.

Content Based Video Retrieval is an important topic of research in the field of Multimedia Video Processing. The scope of content based video retrieval is very wide. It includes pattern recognition, object detection, human action recognition, motion detection, and so on. In our project we have  implemented template matching, a popular Digital Image Processing technique, which is used for matching parts of an image with a template image. We know that a video is nothing but a sequence of image frames which are displayed one after another so quickly that our eyes tend to perceive them as scenes in motion. So by applying Template Matching on every image frame of the video we detect objects of interest contained in the video. The Video Mining System is an interactive application for mining object of interest in the videos. A unique feature of this application is that it allows the user to give video as an input and define the area of interest from the selected key frames. The system first processes the input video to extract key frames from it. The user can then dynamically select the object of interest from the desired key frame. The user should specify the path of the folder containing the videos to be mined called the Video Dataset. The system then performs the mining process by matching the template with the videos. The matched frames are saved in the form of a video.