Detecting Foot-Chases for Police Body-Worn Video

Undergraduate #55
Discipline: Computer Sciences and Information Management
Subcategory: Computer Science & Information Systems

Rafael Aguayo - University of California, San Diego
Co-Author(s): Alejandro Camacho, California State University, Fullerton, CA Qi Yang, University of Southern California, Los Angeles, CA Piyali Mukherjee, Columbia University, New York, NY Hayden Schaeffer, Carnegie Mellon University, Pittsburgh, PA

---

Existing methods to record interactions between the public and police officers are unable to capture the entirety of police-public interactions. In order to provide a comprehensive understanding of these interactions, the Los Angeles Police Department (LAPD) intends to utilize Body-Worn Video (BWV) collected from cameras fastened to their officers. BWV provides a novel means to collect fine-grained information about police-public interactions as well as provide alternative means of evidence.
The research objective is to determine the effectiveness of machine-learning algorithms in identifying specific actions from videos, in particular foot-chases. Our proposed algorithm uses semi-supervised methods such as the detection of point-features and their classification via support-vector machines, which have been implemented using MATLAB (including the Computer Vision and Statistics and Machine Learning toolbox). The detection of point-features includes using methods such as SURF, BRISK, and optical flow, which provide local and global changes in points. Our training dataset consists of 100 training videos (20 foot-chase and 80 non-foot-chase) and a test dataset of 60 LAPD videos (4 foot-chase and 56 non-foot-chase). We achieved results of 91.6% testing accuracy. From the test dataset, we also correctly identified every foot-chase occurrence, which was one of the main goals of this project. From this study, we have concluded that machine-learning techniques are a viable option for identifying foot-chases. Through widespread use of this software, it will contribute to saving valuable resources of the LAPD as their usage of BWV expands and provide a tool for the LAPD to conduct video analytics.

Funder Acknowledgement(s): This study was supported by a grant from the National Science Foundation (NSF grant DMS-0931852).

Faculty Advisor: Hayden Schaeffer,