There is no English translation for this web page.
Performance and Workload Evaluation of Deep Learning Simulation in Disaster Monitor Scenario with Subjective and Physiological Measurements
LOCATION: TEL, Auditorium 3 (20th floor),
Ernst-Reuter-Platz 7, 10587 Berlin
Date/Time: 13.02.2017, 14:15-15:00
SPEAKER: Ziqi Liu (TU Berlin)
In recent years, there is a tendency of more frequent natural disasters because global climate changes over the years. Earthquakes and storms threaten the safety of mankind. What related authorities have been doing to reduce the loss caused by disasters is to collect information from the disaster attacked area and allocate resources to the these areas in need. In typical urgent response agencies, the tactical(Silver) level of disaster response is in charge of monitoring and identifying the disaster events. UAVs with its mobility and flexibility, have been deployed in disaster response scenarios in recent years. When operating UAVs in Silver level, two types of operators are usually needed. One is planner operator to monitor reports of disasters and plan flights of UAVs to identify disaster events. Another is camera operator to supervise the video feeds of UAVs from the ground and tag suspicious disasters in the video. Such kind of deployment is still innovative and many researches are dedicated to find ways to optimize it.
In this study, possibility of replacing camera operator in Silver operation with deep learning algorithm is investigated. Recent scene recognition in deep learning has achieved great progress, it has the advantage of being more scalable regard of information volume being processed and more stable regard of performance compared to human operators. With the assumption that deep learning algorithm can outperform camera operators and meanwhile consume less workload from planner operator when Silver is overloaded with disaster events and multiple UAV camera feeds, a controlled experiment is carried out under simulation. The result of the experiment reveals that deep learning outperforms in accuracy when disaster event load is very high and very low. The other parameters suggest that deep learning can achieve at least as good performance without consuming extra workload from planner operator. In general, the study proves the advantages of applying deep learning algorithm in disaster response. Deep learning can achieve a better performance especially when big disaster happens therefore large amount of disaster related reports flood in the Silver system and when Silver system is falling into idle status on a regular daily bases. At the same time, deep learning can reduce human resources in the disaster responding agencies and will not introduce extra workload to the remaining human operators.