Parts-based detection of AK-47s for forensic video analysis

Abstract

Law enforcement, military personnel, and forensic analysts are increasingly reliant on imaging ystems to perform in a hostile environment and require a robust method to efficiently locate bjects of interest in videos and still images. Current approaches require a full-time operator to monitor a surveillance video or to sift a hard drive for suspicious content. In this thesis, we demonstrate the effectiveness of automated analysis tools to detect AK-47s in images. By training on a large corpus of labeled data, we created Viola-Jones classifiers for detection of whole AK-47s and parts of an AK-47. Parts-based detections were then compared against learned models using support vector machines and multi-layer perceptrons. The results of this research show that parts-based classifiers combined with the above techniques leverage the high recall capability of part detectors and significantly reduce false positives in comparison to both the part and whole object classifiers. Techniques utilized in this thesis facilitate the creation of an automated capability for detecting AK-47s in support of the law enforcement and intelligence communities.