Genetic MRF model optimization for real-time victim detection in search and rescue
Genetic MRF model optimization for real-time victim detection in search and rescue
复制标题
用于搜救中实时受害者检测的遗传 MRF 模型优化
DOI：
10.1109/iros.2007.4399006
复制
发表时间：
2007
期刊：
影响因子：
--
通讯作者：
R. Kümmerle
中科院分区：
文献类型：
--
作者：
A. Kleiner;R. Kümmerle
救援机器人的一个主要目标是在灾难发生后部署一组机器人进行协调的受害者搜索。这就要求机器人实时执行子任务，如受害者检测。人类通过计算成本低廉的技术（如颜色阈值）进行检测，结果会产生大量的误报。利用马尔可夫随机场（mrf）将多个弱分类器的局部证据进行组合，可以提高检测率。然而，磁共振成像中的推理是计算昂贵的。本文提出了一种新的MRF模型构建过程的遗传优化方法。遗传算法确定数据的离线相关邻域关系，然后利用这些邻域关系在运行时从视频流生成高效的MRF模型。实验结果清楚地表明，与基于支持向量机（SVM）的分类器相比，优化后的MRF模型显著降低了误报率。此外，在几乎相同的检测率下，优化后的模型比未优化的模型快5倍。
One primary goal in rescue robotics is to deploy a team of robots for coordinated victim search after a disaster. This requires robots to perform sub- tasks, such as victim detection, in real-time. Human detection by computationally cheap techniques, such as color thresholding, turn out to produce a large number of false-positives. Markov Random Fields (MRFs) can be utilized to combine the local evidence of multiple weak classifiers in order to improve the detection rate. However, inference in MRFs is computational expensive. In this paper we present a novel approach for the genetic optimizing of the building process of MRF models. The genetic algorithm determines offline relevant neighborhood relations with respect to the data, which are then utilized for generating efficient MRF models from video streams during runtime. Experimental results clearly show that compared to a Support Vector Machine (SVM) based classifier, the optimized MRF models significantly reduce the false-positive rate. Furthermore, the optimized models turned out to be up to five times faster then the non-optimized ones at nearly the same detection rate.