Crowd-based velocimetry for surface flows

Abstract

Flow velocity measurement is important in hydrology. Recently, owing to the popularity of sensors and processors, image-based flow velocity measurement methods have become an important research direction. Particle image velocimetry (PIV) is a key example. However, due to the uncertainty of the features, PIV sometimes provides very inaccurate results and always requires customized setups. In this research, we take advantage of the human perception system, that is, the strong abilities related to feature identification and tracking, in order to estimate the surface flow velocity of a river. We developed a method called crowd-based velocimetry (CBV) to incorporate the human perception capacity in the estimation of the flow velocity. CBV includes three main steps: (1) video processing, (2) crowd processing, and (3) statistical processing. We validated CBV by measuring a fast, steady, and uniform river surface flow in an artificial canal. The results show that compared to radar measurements from the center of the flow, CBV measured the surface flow velocity with a deviation ranging between +12.1% and +17.3% from the radar measurement, while PIV resulted in a −1.7% to −24.3% deviation. With rapidly improving mobile devices, CBV allows enormous numbers of people to engage in flow measurement, making CBV more reliable, more efficient, and more economical.

Introduction

Surface flow velocity measurement is an important step in hydrology surveys and disaster prevention. By utilizing the flow velocity distribution, we can calculate the flow rate using a volume flow rate equation [1]. Rating curve estimation corrected by complete flow velocity measurement is a common way to conveniently estimate the flow rate of a river. Lin et al. estimated the flow rate using the surface flow velocity [2]. Costa et al. and Welber et al. validated the feasibility of using surface velocity radar (SVR) to measure the discharge of a river [3], [4], [5]. Flow velocity measurements are classified into two types: contact and non-contact measurements. Contact measurements include mechanical methods, electromagnetic methods, acoustic current-meters, acoustic Doppler current profilers, and other methods using particular devices. These devices require time-consuming and labor-intensive manual operations, making measurements difficult. In contrast, non-contact measurements are a better choice for observers. The continuous-wave radar meter, pulsed-wave radar meter, and image-based methods are common non-contact measurement methods.

We chose to focus our research on an image-based method for the following three reasons. First, the diversity and quantity of physical information obtained from images are greater than that obtained from radar waves. Second, recent abundant research related to image-based methods has made the application easier to develop. Finally, high-quality cameras are becoming more affordable than radar meters.

The image-based flow velocity measurement method, also called image velocimetry, has great potential for application flexibility and technical evolution [6]. Particle image velocimetry (PIV) is the most commonly used image-based flow velocity measurement method. PIV compares the subsequent frames in video clips, identifies similar image features, and calculates the velocity of the flow. It uses a detection window called an interrogation area (IA) to capture the subsequent frames at the same location on the image. It calculates the coefficients of correlation between two IAs and then generates a map of the coefficients. The direction of the feature motion corresponds to the location of the peak from the central point on the map. PIV also indicates the surface flow velocity. We will now review the literature on image velocimetry in two categories: (1) laboratory methods and (2) on-site methods.

In 1991, image-based velocimetry was developed in a controlled laboratory environment and was called particle image velocimetry (PIV) [7]. In PIV, two laser sheets are emitted on a flow with a delay, and simultaneous cameras are used to measure the particles' flow velocity vectors. In the 1990s, PIV techniques were used for flow velocity measurements in many fields, as summarized by Grant in a review [8]. Since spurious vectors and bias errors commonly occur in PIV results, Hart provided an error correction method that improves the spatial resolution and vector yields [9]. Heng utilized an adaptive analysis window and a cross-correlation calculation based on the Hartley transform to accomplish better efficiency and lower error [10]. Subsequently, PIV using different types of tracers [11] or different conditions for hydraulic models [12] was also discussed. These methods are implemented in the laboratory and are the fundamental concepts underlying on-site methods.

In 1998, Fujita et al. discussed large-scale PIV (LSPIV) for measuring flow velocity on-site covering an area of 4–45,000 square meters with three applications [13]. Following this, many studies implemented LSPIV for estimating river flow, including measurements for large rivers [14], [15], [16], [17] and small rivers or lakes [18], [19], [20]. However, LSPIV still has certain limitations. Here, we will consider two examples. First, PIV analysis requires a high computation time. Second, environmental uncertainties such as natural light and weather conditions cause noise on images. To address the first issue, Fujita et al. developed space-time image velocimetry, assuming that the flow velocity direction is downstream. Although it improves the computational efficiency, information regarding the two-dimensional flow direction is lacking [21]. To address the second issue, some studies integrated cameras with different devices to improve the PIV performance in a varying environment. To eliminate noise due to natural light, Li et al. captured images using multi-channel charge-coupled device (CCD) cameras [22]. Zhang et al. enhanced patterns on a river surface using a near-infrared radiation (NIR) filter [23], and Wang et al. used a NIR filter on a balloon that captured images under extreme weather conditions [24]. Those techniques applied additional equipment to reduce the impact of environmental uncertainties.

According to the description of using PIV measurements in open-channel flows [25] and our experiences in PIV measurements, PIV may not be applied on-site for two reasons. First, parameters such as the IA and step sizes are difficult to determine. Second, it is difficult to identify the result in the case of a significant temporally varying flow velocity since significant features are randomly shaped and occur randomly and in a scattered manner. PIV inspects all temporal and spatial domains of the region of interest (ROI) in a video. The features, including velocity-relevant ones and velocity-irrelevant ones, are all considered in PIV analysis. Although the relevant features are applied to measure the flow velocity, the irrelevant features, such as moving shadows and swaying branches and leaves, may interfere with the flow velocity measurement. Moreover, the irrelevant features make the results vary considerably if such features constitute the major particles of an image. However, a pre-attentive process involving the human visual perception system can be used to filter and determine significant, or velocity-relevant, features in a video [26]. Therefore, we took advantage of the human visual perception system to measure the flow velocity. However, due to the large amount of image data involved, visual inspection by a single person would be difficult, and so we used a group of people (crowd) to measure the flow velocity. Some studies have discussed the application of crowdsourcing on hydrology measurement [27], [28]. Regarding crowd-sourced image data for flow velocity estimation [29], [30], [31], our research goal is to develop a method to estimate the flow velocity using crowd intelligence without expert experience and iterative parameter settings, and to create a user interface to realize this purpose. Hence, based on our method, the public can be engaged in flow measurement with a mobile device.