Amrita Center for Wireless Networks and Applications (Amrita WNA) was provided with an opportunity to present a poster on Urban Flood Monitoring using Crowd Sourcing and Computer Vision at MobiSys 2017, the 15th ACM International Conference on Mobile Systems, Applications and Services, held at Niagara Falls, NY, USA, on June 20, 2017.
The poster titled, “Flood Monitoring Using Computer Vision”, was authored by Bhavana B. Nair, Research Associate Trainee at Amrita WNA and Sethuraman N. Rao, Associate Professor at Amrita WNA. The project uses computer vision on crowd-sourced images that are captured with a smart phone for effective and timely urban flood relief and management. The data can also be used to assess the effectiveness of preventive measures taken in the past and to plan remedial measures in the future.
The methodologies involved are the input of jpeg images involving humans in flood scenes. The human face is then detected using Dlip face detection algorithm. A deep learning framework known as Caffe, developed by Berkeley AI Research (BAIR), is used to segment out humans and also to classify the human faces based on gender. Next the depth of water is estimated. The average human height is taken as- Male: 5’ 6” and Female: 5’ 0”. Human height in the image is extrapolated from the height of the face obtained from the face detection algorithm based on the Golden Ratio Phi ( = 1.618). The ratio of body height to face height is taken as "Phi ^ 4 = 6.854". The human segmented image is used to find the water line within the human frame. The water depth is estimated based on the locations of the water line and the human feet in the human frame and the value of average male/female human height.
The challenges faced in this project are that human face and gender detection algorithm may fail with low resolution images. Lighting conditions and occlusion can also affect the results. Human reflection in water is considered as part of human segmentation by the deep learning algorithm. This will affect the accuracy of results. Future works include: water segmentation to find the water line; further refinement of the reference human height based on detection of age, ethnicity, etc. and water depth estimation based on reference height of other objects such as vehicles, buildings, electric posts, etc.,in the image.