Computer Vision projects
Justin Hess
Jordan Helderman
This project uses various algorithms which utilize temporal derivatives in order to detect moving people using a video recorded on a stationary camera. The zip file is provided which contains all the images that combined make a video. Using a one dimensional, temporal derivative filtering mask, we can detect drastic shifts in the grayscale intensity of pixels on a stationary background when a moving object passes them along the time spectrum. By looking at the large temporal gradients in the pixels where a moving object passed, we can detect and track the motion of the moving objects. To accomplish this, we created a binary mask by thresholding the absolute value of the temporal derivatives and combined the result with the mask to highlight the pixels around the moving objects. Several 2D spatial smoothing algorithms were also used including two box filters and a gaussian filter with varying the sigma value.
This projects utilizes various scripts and function to create a projective transformation between two image spaces, otherwise known as image mosaicing. Two images are warped into the same image space to create an image that is the union of both images. This is accomplished by detecting corner points in a image, correlating those points with corresponding points in a similar image, using those points to estimate a homography between the two image spaces, and warping the images into the same image spaces. A Harris corner detection function in MATLAB was implemented and used to apply normalized cross correlation to find corresponding corner features between the given images. RANSAC was also used to estimate the homography in a fashion robust against incorrect correspondences. Finally, images are warped together using reverse image warping.
In computer vision, there is sometimes a need for understanding the apparent displacement of an object of interest between several images of a scene, whether it is due to motion of the objects in the scene or the motion of the camera. This information is known as the optical flow, and it is useful in many video based tracking applications. In this project, an algorithm for estimating the optical flow, the Lucas-Kanade algorithm, is implemented. This algorithm can be used to estimate the flow vectors in a benchmark set of images.
Target tracking is an important topic in computer vision, having application in automatic surveillance applications and, in general, in video processing applications. One challenge in target tracking is detecting and recovery from occlusion. In this project, various algorithms for occlusion detection and recovery are implemented. Peak-to-sidelobe ratio is used as a detection statistic for occlusion detection and two methods for occlusion recovery are experimented, which both predict the location of the occluded object based on amodel for the dynamics of the object. The first method is a simple constant velocity model, and the other is an adaptive model that automatically estimates the complexity of the motion with the Hankel matrix and predicts the position of the occluded object via a linear predictor.