Detecting and tracking people in real-time

Abstract

The problem of detecting and tracking people in images and video has been the subject of a great deal of research, but remains a challenging task. Being able to detect and track people would have an impact in a number of fields, such as driverless vehicles, automated surveillance, and human-computer interaction. The difficulties that must be overcome include coping with variations in appearance between different people, changes in lighting, and the ability to detect people across multiple scales. As well as having high accuracy, it is desirable for a technique to evaluate an image with low latency between receiving the image and producing a result. This thesis explores methods for detecting and tracking people in images and video. Techniques are implemented on a desktop computer, with an emphasis on low latency. The problem of detection is examined first. The well established integral channel features detector is introduced and reimplemented, and various novelties are implemented in regards to the features used by the detector. Results are given to quantify the accuracy and the speed of the developed detectors on the INRIA person dataset. The method is further extended by examining the prospect of using multiple classifiers in conjunction. It is shown that using a classifier with a version of the same classifier reflected in the vertical axis can improve performance. A novel method for clustering images of people to find modes of appearance is also presented. This involves using boosting classifiers to map a set of images to vectors, to which K-means clustering is applied. Boosting classifiers are then trained on these clustered datasets to create sets of multiple classifiers, and it is demonstrated that these sets of classifiers can be evaluated on images with only a small increase in the running time over single classifiers. The problem of single target tracking is addressed using the mean shift algorithm. Mean shift tracking works by finding the best colour match for a target from frame to frame. A novel form of mean shift tracking through scale is developed, and the problem of multiple target tracking is addressed by using boosting classifiers in conjunction with Kalman filters. Tests are carried out on the CAVIAR dataset, which gives representative examples of surveillance scenarios, to show the performance of the proposed approaches.