Repository hosted by TU Delft Library

Home · Contact · About · Disclaimer ·

Quantifying the uncertainty of event detection in full motion video

Publication files not online:

Author: Neumann, N.M.P. · Knegjens, R. · Hollander, R. Den · Oggero, S. · Burghouts, G.J. · Broek, S.P. Van Den
Publisher: SPIE
Source:Counterterrorism, Crime Fighting, Forensics, and Surveillance Technologies II 2018, 10 September 2018 through 11 September 2018, Bouma, H.Prabhu, R.Stokes, R.J.Yitzhaky, Y., Proceedings of SPIE - The International Society for Optical Engineering, 10802
Identifier: 844195
Article number: 108020E
Keywords: Detection · Full Motion Video · Machine Learning · Tracking · Uncertainty · Artificial intelligence · Error detection · Learning systems · Monitoring · Motion analysis · Pipelines · Security systems · Surface discharges · Terrorism · Video recording · Airborne platforms · Data-driven approach · Detection and tracking · Detection characteristics · Full motion video · Time constraints · Uncertainty · Video-surveillance applications · Chemical detection


Algorithms for the detection and tracking of objects can be combined into a system that automatically extracts relevant events from a large amount of video data. Such a system , can be particularly useful in video surveillance applications, notably to support analysts in retrieving information from hours of video while working under strict time constraints. Such data pipelines entail all sort of uncertainties, however, which can lead to erroneous detections being presented to the analyst. In this paper we present a novel method to attribute a confidence of correct detection to the output of a computer vision data pipeline. The method relies on a datadriven approach. A machine learning-based classifier is built to separate correct from erroneous detections. It is trained on features extracted from the pipeline; The features relate to both raw data properties, such as image quality, and to video content properties, such as detection characteristics. The validation of the results is done using two full motion video datasets from airborne platforms; the first being of the same type as the training set, the second being of a different type . We conclude that the result of this classifier could be used to build a confidence of correct detection, separating the True Positives from the False Positives. This confidence can furthermore be used to prioritize the detections in order of reliability. This study concludes by identifying additional work measures needed to improve the robustness of the method. © 2018 SPIE.