Year: 2016
Type: article
Abstract: Spectator Performer Space (SPS) is a frequently occurring crowd dynamics, composed of one or more central performers, and a peripheral crowd of spectators. Analysis of videos in this space is often complicated due to occlusion and high density of people. Although there are many video analysis approaches, they are targeted for individual actors or low-density crowd and hence are not suitable for SPS videos. In this work, we present two trajectory-based features: Histogram of Trajectories (HoT) and Histogram of Trajectory Clusters (HoTC) to analyze SPS videos. HoT is calculated from the distribution of length and orientation of motion trajectories in a video. For HoTC, we compute the features derived from the motion trajectory clusters in the videos. So, HoTC characterizes different spatial region which may contain different action categories, inside a video. We have extended DBSCAN, a well-known clustering algorithm, to cluster short trajectories, common in SPS videos. The derived features are then used to classify the SPS videos based on their activities. In addition to using NaïveBayes and support vector machines (SVM), we have experimented with ensemble based classifiers and a deep learning approach using the videos directly for training. The efficacy of our algorithms is demonstrated using a dataset consisting of 4000 real life videos each from spectator and performer spaces. The classification accuracies for spectator videos (HoT: 87%; HoTC: 92%) and performer videos (HoT: 91%; HoTC: 90%) show that our approach out-performs the state of the art techniques based on deep learning. Advisor: Ashok Samal
Source: Lincoln (University of Nebraska)
Author: Anish Timsina
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2478597682