Abstract

Deep convolutional neural networks (CNNs) have proven very effective for many vision benchmarks in object detection and classification tasks. However, the computational complexity and object resolution requirements of CNNs limit their applicability in wide-view video surveillance settings where objects are small. This paper presents a CNN surveillance pipeline for vessel localization and classification in maritime video. The proposed pipeline is build upon the GPU implementation of Fast-R-CNN with three main steps:(1) Vessel filtering and regions proposal using low-cost weak object detectors based on hand-engineered features. (2) Deep CNN features of the candidates regions are computed with one feed-forward pass from the high-level layer of a fine-tuned VGG16 network. (3) Fine-grained classification is performed using CNN features and a support vector machine classifier with linear kernel for object verification. The performance of the proposed pipeline is compared with other popular CNN architectures with respect to detection accuracy and evaluation speed. The proposed approach mAP of 61.10% was the comparable with Fast-R-CNN but with a 10× speed up (on the order of Faster-R-CNN) on the new Annapolis Maritime Surveillance Dataset.


Original document

The different versions of the original document can be found in:

http://dx.doi.org/10.1109/avss.2016.7738076 under the license cc0
https://dblp.uni-trier.de/db/conf/avss/avss2016.html#BousetouaneM16,
https://ieeexplore.ieee.org/document/7738076,
https://academic.microsoft.com/#/detail/2551589584
Back to Top

Document information

Published on 01/01/2016

Volume 2016, 2016
DOI: 10.1109/avss.2016.7738076
Licence: Other

Document Score

0

Views 2
Recommendations 0

Share this document

Keywords

claim authorship

Are you one of the authors of this document?