Detecting Water in Visual Image Streams from UAV with Flight Restrictions

Abstract

Water surface identification is useful for identifying water retention areas, water leaks, puddles, and drone safety systems. For computer vision approaches, water surface identification is a complex task since water surfaces do not possess exact shape or colour. This research focuses on identifying water surfaces from drone images under restricted heights and camera orientations. For the identification process, machine learning and computer vision approaches were implemented and evaluated. In this research, three aspects were addressed, as (1) improving the accuracy of FCN-RAU [1] on identifying reflection property on water by using recent advancements in deep learning based segmentation techniques, (2) detecting water surfaces when hovering based on the ripples generated, (3) detecting water surfaces from static altitude movement (SAM) by using the effect of uniform texture when compared to the ground. In this research, a new dataset was introduced as the UAV-W dataset. It contains images and videos of water surfaces taken by drones at different altitudes and camera orientations. UAV-W’s DRONE subset contain 119 front facing drone images and their manual annotations. Puddle-1000 [1] and UAV-W datasets were used for experiments. When a water surface is viewed from a distance, the water surface acts as a mirror which reflects the surrounding environment [1]. In this research, UNet [2] – a recent advancement in visual segmentation with deep convolutional neural networks was investigated for segmenting such reflections. It provides a network and a training strategy that relies on data augmentation to use the available annotated samples more efficiently. The gist for the Reflection Attention Unit(RAU) is reflections on a water surface of far away objects appear to be in a symmetrical vertical line [1]. UNet improved performance of the state of the art method of FCN – RAU [1] in water detection. For the Puddle-1000 dataset [1], UNet [2] achieved F1-scores of 79.42% on on-road dataset (ONR), 86.28% on off-road dataset (OFR), and 83.81% on both on-road and off-road datasets (BOTH) combined. This is an improvement of 9.31% on ONR, 4.61% on OFR, and 6.90% on BOTH datasets in F1-score. On drone imagery presented here as UAV-W’s DRONE subset, UNet produced 98.87% of accuracy, 80.43% of recall, 81.77% of precision, and 81.09% of F1-score. When a drone hovers close to a water surface, circular-like wave formation is observed. Hough Circles Algorithm [3] was used to detect water surfaces, by identifying circular-like ripples on the water while Dense Optical Flow Algorithm [4] was used to identify water surfaces by capturing the ripple movement. At altitudes between 2-5m and 1-5m, Hough Circles and Dense Optical Flow algorithms detected water surfaces, respectively. When a drone moves at higher altitudes (10-50 m), water surfaces in the video stream are visible as uniform texture regions compared to the surrounding environments. Dense Optical Flow method was successful in identifying water surfaces by using this method at altitudes 10-50 m. It is concluded that UNet trained on Puddle-1000 [1] performance better than FCN - RAU [1]. This can be successfully integrated with a drone vision system, in identifying water surfaces. Optical flow approach works for both hovering (1-5 m) and SAM (10-50 m) when identifying water surfaces.