Real-time traffic congestion detection based on frame difference function and virtual loop

Abstract

Traffic congestion has become one of the major concerns of policy-makers in modern metropolises. Accurate real-time traffic congestion alert is of great importance for alleviating congestion. In this paper, we propose a fast, unsupervised, video-based approach using average frame difference function (AFDF) and virtual loop to identify real-time traffic congestion. This novel method calculates the frame differences of specified order in a loop area to determine the speed of the vehicles. The proposed method has been integrated in the intelligent transport system (ITS) of Wuhan City, China for testing, and results show that the method is efficient and robust for real-time traffic congestion detection.. Introduction As the biggest city in Central China with more than 1.7 million vehicles registered, Wuhan is suffering from severe traffic congestion problems like many of the other big cities of the developing nation. The demands for traffic congestion alert is urgent for traffic management. Besides, accurate and real-time traffic information including the traffic congestion status is one of the fundamental parts of ITS. In this paper, we focus only on the status of traffic congestion, which means that the output of the proposed method is whether the traffic is congested or not in a certain lane section. The algorithm can be easily applied to multi-lanes applications. Many researchers have dedicated their effort to construct the congestion detection algorithm in recent years. Vehicle-to-infrastructure (V2I) methods such as [1] and vehicle-to-vehicle (V2V) methods such as [2] both require the on board units or sensors which make it less feasible and efficient when huge numbers of vehicles are taken into account. Compared with V2I or V2V, video-based approaches are less expensive and is easy to maintain. Most of the existing video-based approaches [3-6] have not kept a good balance between accuracy and real-time performance. In this paper, we combine the concept of virtual loop which is inspired by magnetic loop sensor [7] and a modified frame difference method to detect traffic congestion. This method only requires a day-night infrared camera that is able to work 24 hours a day and a computer that has the access to the camera. It is rather cheap and is easy to maintain. And it only engages small amount of pixel-wise calculation which ensures the system’s ability to run with a great speed. Tests show that the method responses quickly and the accuracy of congestion detection is satisfactorily high. Virtual loop and average frame difference function The average frame difference function (AFDF). It stands to reason that traffic congestion can be interpreted as a state that little difference exists between adjacent frames, which makes frame difference a potential criteria for traffic congestion identification. Instead of calculating the whole frame, a user-defined virtual loop is used for calculation, which is shown in Fig. 1. The camera is mounted up above the one-way road, and at the bottom of each lane a virtual loop is defined. The red rectangles are virtual loops, and yellow envelops are the corresponding region of interest (ROIs) of three different lanes. When the traffic status of each of the lanes is known, then the Joint International Mechanical, Electronic and Information Technology Conference (JIMET 2015) © 2015. The authors Published by Atlantis Press 670 traffic status of the whole road can be determined. Thus the point of detecting congestion is to determine the state of each single lane which can be inferred from the changes of AFDF. The nth order AFDF is defined as follows: f(i) = mean{[I(i) − I(i − n)]loop} (1) I(i) is the ith frame of the video. The subscript loop means the ROI regions of the calculation. The region out of the loop will not be considered. mean is the average operator. We calculate the f(i) for all the frames behind the nth frame. Fig. 2 is obtained by calculating the AFDF in the region of the middle loop after every 2, 3, 4, 5, 6 frames with respect to frame time. The source frames are converted to gray images for calculation and essential preprocessing actions such as image smoothing and image de-nosing have been applied to the frames. The source video is about 50s long. The video starts with a normal traffic state and then falls into a traffic jam till the end. As the video shows, three vehicles intrude the area of the virtual loop and the former two leave immediately except that the speed of the first vehicle is much bigger than the second one. The third vehicle leaves at about 25 seconds later and another one drives in in one or two seconds. Then almost no vehicles move. Figure 1. Virtual loops and ROIs Figure 2. AFDF of different orders As Fig. 2 shows, the whole actual process has a strong connection with the AFDF especially the 3rd and the higher order AFDFs. The basic point is, higher values of AFDF indicate good traffic conditions and the values below a certain level correspond to congestion or free road. There is no denying the AFDF is a good indicator of whether the vehicles are floating. And after a thoroughgoing study of the results shown in Fig 2, there are some basic verdicts about the AFDF that can be concluded: • The move of vehicles produces an impulse-like change in the AFDF in all orders while the period of traffic jam corresponds to the flat area of the function. • First order AFDF is sensitive to fast moving vehicles but is very dull to slow ones. The boundary between congestion and free traffic is sharply clear. But mistakes exist when vehicles move a very small distance in congested traffic. • The response of higher order AFDF to slow changes in the video is stronger than that of the first orders’. Higher order brings higher detail resolution. We can even find the exact time that a vehicle drives into or out of the virtual loop from the sixth order AFDF. • The fourth order AFDF is so different from the other orders’. It has strong response to fast moving vehicles than the lower-orders’ and preserves a relative high detail resolution.

Original document

The different versions of the original document can be found in:

• http://dx.doi.org/10.2991/jimet-15.2015.125 under the license https://creativecommons.org/licenses/by-nc

• https://download.atlantis-press.com/article/25843841.pdf

• https://www.atlantis-press.com/proceedings/jimet-15/25843841,

https://academic.microsoft.com/#/detail/2254739418 under the license cc-by-nc