Drone-surveillance for search and rescue in natural disaster
Introduction
Search and rescue (SAR) have been a human-intensive task so far, but recent technological advancements can make it autonomous. Using drone surveillance with a recent computer vision technology can increase the number of humans saved at the time of disaster. However, recent activities of using the drone have mixed reaction, but using a drone to save someone’s life is novel and a great cause. Recently, drones are being used more for SAR and providing excellent support in those operations. These days, police and fire departments have also adopted drones and collaborated with local SAR teams for time-sensitive rescue operations. In January of 2019, a SAR team in Snowy Canyon State Park, Utah, used a drone to help rescue a hiker trapped on a ledge at night. The hiker was 60 years old, and SAR first found that he was trapped from other hikers who heard the man calling out for help [1]. Also, a rescue event in Texas, drone was used to find an 88-year-old missing man [1]. In another story of using the drone for SAR where it is used to find two cousins trapped on a mountainside in Iceland [1]. All these incidents show the capability and importance of the drone in the field of SAR however, these operation are performed manually for finding the person which need an automation to apply at bigger scale. These autonomous drone with on-device video analysis capability for saving life motivates us to develop a novel and dedicated system for autonomous searching of people who are stuck and required rescue.
The idea of drone-surveillance for SAR is to use the drone for scanning the affected area with the help of camera, and model deployed on the drone itself for identifying the exact places where help is required. An example of automated surveillance and search operation is shown in Fig. 1. In this figure, after the identification of humans location, GPS location of human can be sent to the rescue team for the fast and productive rescue. The recent success of deep-learning approaches for object detection and action recognition motivates us to apply it in the drone-surveillance. The essential part of a deep-learning approach is that a significant amount of data is needed for training. Since, most of the dataset available in the literature are for ground-level surveillance such as UCF [2], which is not useful for training deep-learning model of aerial surveillance. Hence, it is our primary objective to develop a dataset of aerial action recognition for SAR. In addition, deep-learning models uses these dataset for training in different type of task such as classification and localization. Deep-learning models used for these task can automatically extracts the feature. Out of all other neural networks used for classification or localization, convolution neural networks (CNN) suits more for image-based feature extraction. In CNN, each layer uses a convolution filter for feature extraction. An example of two different type of such task is represented in Fig. 2. In this, detection is a combination of classification and localization. The classification problem of images is mainly to classify the image into a different category (labels), while the objective of detection is to identify the label of the object as well as to determine the exact position of classified labels in that image.
As dataset plays a crucial role in the performance of model, this paper proposes a unique dataset of aerial action recognition for SAR. Also, in the aerial surveillance, since, human appears very small and existing algorithms are not able to identify the action performed by them, this paper also proposes a modified action detection model for aerial action detection. The main contribution of the paper is as follows:
- •
In order to develop any application using deep-learning, the primary requirement is availability of labelled dataset. But for automated search of human using drone surveillance there is no such dataset available in literature. Therefore, in this paper, we have proposed a novel dataset to search humans in rescue for disaster management application.
- •
Proposed dataset is annotated for two different set of action and is available in the form two action dataset and six action dataset for SAR.
- •
In addition, an experimental analysis of deep learning object detection models such as Faster R-CNN, R-FCN, and single shot detection (SSD) applied to existing aerial action detection dataset [3] and proposed dataset, has been presented in the paper. Moreover, an modified SSD has also been proposed for better performance in aerial surveillance.
Section snippets
Related work
Here, we briefly introduce the current work in the field of dataset and the application of model for aerial human action recognition.
Dataset development
This section describes the dataset collection, type of action recorded, pre-processing, and the usefulness of dataset for vision-related real-life applications.
Performance evaluation metrics
mAP and IOU are the standard COCO evaluation parameters used to the evaluate object detection models. Hence, to compare our proposed model with the state-of-the-art models in literature, these parameters are suitable. The details of these parameters are discussed in this section. Also, as precision and recall are two basic parameters on which all these evaluation parameters depends, a brief description of these parameters are also given in this section.
Proposed framework
Proposed architecture with the developed dataset for action detection in drone surveillance, can be used to identify the situations where humans are asking for help. As shown in Fig. 1, on board autonomous analysis of drone images can quickly find humans stuck in the disaster prone area. Proposed dataset is generalized and have enough variation to be used for the automation of such application. In addition, the architecture of proposed model for detecting the action is shown in Fig. 6. Proposed
Experimental setup
Experiments were performed on the NVIDIA DGX-1 V100 supercomputer having 7.8 TFLOP/s for FP64 computation power. To recognize human action, state-of-art object detection was applied to the proposed dataset. In addition to this, experiments were performed on publicly available Okutama dataset as well. Both datasets have frames that contain multiple people performing different actions simultaneously. As the image size of both datasets is equal, i.e., 1920 * 1080 pixels, the results of the various
Results and analysis
In this section, we have discussed the results in detail based on the availability of visual and statistical results for human detection and action recognition.
Table 5 shows the performance of deep learning object detection models applied on publicly available Okutama dataset. The performance is evaluated on a standard coco evaluation metric (mAP). Our result shows that faster R-CNN is performing comparatively better on this dataset. In addition to this, Table 6 shows the results of models
Conclusion
In this paper, we have proposed a drone dataset for human action recognition. This dataset can also be used for human detection and other such task for different surveillance applications. Proposed dataset has a rich amount of variety in terms of colour, height, actor, and background. This variation makes it generalized for proposed dataset to be used for various applications. In addition, as our primary objective is to provide the support for SAR using drone surveillance, we have presented an
CRediT authorship contribution statement
Balmukund Mishra: Conceptualization, Methodology, Writing - original draft. Deepak Garg: Data curation, Investigation, Supervision, Writing - original draft, Project administration. Pratik Narang: Visualization, Investigation, Writing - original draft. Vipul Mishra: Data curation, Formal analysis, Supervision, Writing - original draft, Visualization.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (39)
- et al.
Recent methods and databases in vision-based hand gesture recognition: a review
Comput. Vis. Image Underst.
(2015) Drones in search and rescue: 5 stories showcasing ways search and rescue uses drones to save lives
(2019)- et al.
UCF101: A dataset of 101 human actions classes from videos in the wild
(2012) - Mohammadamin Barekatain, Miquel Martí, Hsueh-Fu Shih, Samuel Murray, Kotaro Nakayama, Yutaka Matsuo, Helmut Prendinger,...
- Joao Carreira, Andrew Zisserman, Quo vadis, action recognition, a new model and the kinetics dataset, in: Proceedings...
- et al.
Review of action recognition and detection methods
(2016) - et al.
Convolutional neural networks for aerial multi-label pedestrian detection
- Asanka G. Perera, Yee Wei Law, Javaan Chahl, UAV-GESTURE: a dataset for UAV control and gesture recognition, in:...
- et al.
Selective search for object recognition
Int. J. Comput. Vis.
(2013) - et al.
SPP-NEt: Deep absolute pose regression with synthetic views
(2017)
Faster r-cnn: Towards real-time object detection with region proposal networks
Saliency detection with recurrent fully convolutional networks
Ssd: Single shot multibox detector
Object detection with discriminatively trained part-based models
IEEE Trans. Pattern Anal. Mach. Intell.
Fast deep vehicle detection in aerial images
Vehicle detection from high-resolution aerial images using spatial pyramid pooling-based deep convolutional neural networks
Multimedia Tools Appl.
Comprehensive analysis of deep learning based vehicle detection in aerial images
IEEE Trans. Circuits Syst. Video Technol.
Cited by (222)
Exploring innovative techniques for damage control during natural disasters
2024, Journal of Safety Science and ResilienceA Machine learning approach for Post-Disaster data curation
2024, Advanced Engineering InformaticsGlobal strong convexity and characterization of critical points of time-of-arrival-based source localization
2024, Computational Geometry: Theory and ApplicationsMulti-objective reinforcement learning for autonomous drone navigation in urban areas with wind zones
2024, Automation in ConstructionExploring artificial intelligence for applications of drones in forest ecology and management
2024, Forest Ecology and Management