People surveillance using video object tracking

Abstract

Background: Pedestrian tracking is a sophisticated topic in the realm of video object tracking. There are many approaches towards the problem of people tracking. The problem can however be generally divided into two parts. "Pedestrian detection" and "Data association". The task for pedestrian detection is to detect appearance of people in a frame, while Data association associate the detections from consecutive frames together to form a tracklet. Problem: One of the main difficulties of pedestrian tracking is to handle occlusion. Occlusion can happen when camera view of pedestrians is being blocked from some non-pedestrian objects. For instance, buildings, vehicle, etc. Importance : Pedestrian tracking is useful in many scenarios. From security system, to automated vehicle. Pedestrian tracking system can perform monitoring, surveillance on pedestrian without inspection from a human, and with a higher stability. Conducted Study: Machine learning and image processing are the main research domains of the project, yet they are closely related. To tackle the problem of people detection, 2-class classification methods have been looking on. For instance, SVM (support vector machine). Feature extraction of an object is also needed. HOG (Histogram of Oriented Gradient) descriptor is being studied in this project. Other method like background subtraction is also being study. For data association part, Hungarian Algorithm and Kalman filter are the main tools being studied, while other method like Mean-Shift is also being studied. The implementation is based on OpenCV library in C++ on window platform. Main results: The implementation successfully performs pedestrian detection on frames of a video, low false positive rate and good true positive rate can be observed. The association of the detections to form tracklets also shows success. It shows some robustness in handling occlusion. The implementation of the data association however still suffers from complex situation when people crossing each other in a crowded scene, the tracks may be wrongly assigned. Iit also suffers from low frame rate if the video resolution is high. Conclusion: Pedestrian tracking is still a very open topic, different approach will have different advantages and drawbacks in different scenarios. In this project, we focus on "Pedestrian detection" and "Data association approach and expose it advantages and weaknesses.