Pedestrian Detection and Tracking in Surveillance Video

Pedestrian detection and tracking has many important applications in the security industry, pedestrian demographic analysis, and intelligent transportation system (ITS). In this project, we will develop a stable pedestrian detection and tracking algorithm. The Town Centre video frames and the hand annotated ground truth published by the University of Oxford are used as a benchmark. In experiment 1, we used Dalal and Triggs (2005) Support Vector Machines (SVM) classifier to detect pedestrians. In experiment 2, we trained our own cascade of boosted classifiers with Histogram of Oriented Gradients (HOG) feature for detection. Using Daimler training samples and INRIA training samples, we have trained two different detectors to perform pedestrian detection. The detector trained with Daimler training samples has outperformed the detector trained with INRIA training samples. For both experiments, the raw detection results are passed to the tracker. Our tracker uses Kalman filter to estimate the location of the pedestrians based on their track history. For data association, we employed the Hungarian algorithm. Overall, experiment 2 shows a more promising result as compared to experiment 1. Using raw detections from Daimler detector, our multiple object tracking accuracy (MOTA) value in experiment 2 had surpassed Benfold and Reid (2011) MOTA value by approximately 1%. However, it was observed that our algorithm suffers from a high number of misses due to occlusion. This is a common problem especially in crowded or semi-crowded environment. Thus to improve the detection or tracking results, one can opt to use a part-based detector instead of a full body detector to estimate the location of the pedestrians.