Contour-based Human Detection with Foreground Distribution Trend Filtering and Tracking

Abstract

In video surveillance detecting human in crowded environment is profitable but challenging Based on contour-based human detection we add two new methods: foreground distribution trend filter and tracking Foreground distribution trend filter deletes those distinguishable false alarms by analyzing foreground distribution trend along x-axis of false alarms Tracking is to record the trajectory a human pass We use this feature to recover those missing detected humans by interpolating positions and consequently increase recall We also make a hypothesis that false alarms may only occur in few frames so we can delete them to increase precision Experimental results show that our proposed methods can improve precision and recall

Date of Award	2014 Jul 14
Original language	English
Supervisor	Pei-Yin Chen (Supervisor)