Video-Centric Applications

Several previous projects developed and experimented with a wide range of image and video processing applications. These include object tracking, hyper-spectral applications, PET image reconstructions, image transforms (e.g., DFT, DCT, wavelet transforms, rotation), image enhancement (e.g., magnification, median filtering), compression (e.g., vector quantization, motion estimation, MPEG compression), and image analysis (e.g., morphological processing, region autofocus, and K-means classification).

Applications
Description
Tracking
Real-time Pedestrian Tracking Tracks people who are moving in a scene (e.g., walking, running, or riding bikes along a sidewalk, indoor corridor, crosswalk, etc.) and computes statistics of pedestrian traffic flow.
Ball Tracking in Sports Video Tracks a moving ball and predicts its future trajectory. Circularity and eccentricity metrics are used to identify candidate balls and path prediction is performed using classic curve fitting.
Image Transforms
Discrete Fourier Transform Transforms an image from the spatial domain to the frequency domain (and vice-versa).
Discrete Cosine Transform Exploits the spatial redundancy inherent in image data, and it is a fundamental component of image compression standards.
Discrete Wavelet Transform Used in image transmission and compression standards.
Image Rotation Two-step rotation operator to rotate the image in the focal plane by any specified angle.
Image Enhancement
Intensity Level Slicing Binarizes the input image, to reduce its bit depth to 2 bpp.
Convolution Performs different filtering operations, such as shadowing, smoothing, and edge detection.
Magnification Performed in digital zooms for cameras and camcorders to enlarge some portion of the image by a given factor.
Median Filtering Removes binary noise from an image while preserving spatial resolution.
Image and Video Compression
Quantization Selects a discrete number of symbols onto which to map the amplitudes according to the level of information content (used in compression standards).
Vector Quantization Compresses and quantizes collections of input data, as an alternative to scalar quantization, by mapping small clusters of input data into a predefined set of symbols.
Entropy Coding Maps the sequence of symbols produced by a quantization process into binary words (used in compression standards).
JPEG Compression Still-image compression standard.
Motion Estimation Removes temporal redundancies between video frames. It is a core building block in several video compression standards, such as MPEG, H.263 and similar.
MPEG Compression MPEG-1 compression standard.
Image Analysis
Morphological Processing Performs feature extraction and segmentation of binary images.
Region Representation Generates a more efficient representation of the original image using a quadrant tree algorithm, to quickly identify regions of interest.
Region Autofocus Isolates a small region of interest within a given image based on some selected features, and it provides a magnified version of this image for further analysis.
K-means Classification Segments large images into specific objects or areas of interest. Typically, this technique is useful in search and rescue operations where large areas must be scanned for specific objects.
Color Imaging
VSobel Extracts color edge information from an image through a Sobel operator that accounts for local changes in both luminance and chrominance components.
SMF Removes impulse noise from an image by replacing each color component with a median value in a 3 × 3 window that is stepped across the entire image. The three resulting images are then combined to produce a final output image.
VMF Suppresses impulse noise from an image through a vector approach that is performed simultaneously on three color components (i.e., Y, Cb, and Cr).
Color Vector Quantization Compresses and quantizes collections of input data by mapping k-dimensional vectors in vector space Rk onto a finite set of vectors. A full search vector quantization using both luminance and chrominance components is used to find the best match in terms of the chosen cost function.
FSVBMA Removes temporal redundancies between video frames in MPEG/H.26L video applications. A full search block-matching algorithm using both luminance and chrominance components is used to find one motion vector for all components.
Chromakey Combines foreground and/or background frames into a final image (e.g., television weather programs).

Publications: