Traffic Speed Measurement using Spatio-Temporal Model and Frequency Domain Analysis Arucha Rungchokanun 1 , Vutipong Areekul 2 , Supakorn Siddhichai 3 , Hiroaki Kunieda 4 1,2 Kasetsart Signal & Image Processing Laboratory (KSIP Lab), Department of Electrical Engineering, Kasetsart University, Thailand, 10900 1 E-mail: kayatao@gmail.com, 2 E-mail: vutipong.a@ku.ac.th 3 Image Technology Laboratory, National Electronics and Computer Technology Center, Pathumthani, Thailand, 12120 E-mail: supakorn.siddhichai@nectec.or.th 4 Department of Communication and Integrated Systems, Tokyo Institute of Technology, S3-65, 2-12-1 Ookayama, Meguro-ku, Tokyo, 152-8550, Japan E-mail: kunieda@vlsi.ss.titech.ac.jp Abstract— In traffic monitoring applications, traffic speed is an important parameter of traffic management. The method for traffic speed measurement using video based on Spatio- Temporal (ST) model and frequency domain analysis is proposed in this paper. This method is designed to be able to measure the traffic speed in every pattern of road. The novel of proposed method is unfolding the edge information on ST model and analyse them in frequency domain to determine traffic speed. The proposed method is evaluated by compare with actual speed. The traffic speed is measured accurately 97.1%. Keywords— Traffic Speed, ST Model, Edge Feature, Frequency Domain, Perspective Effect, Uncalibrated Camera I. INTRODUCTION Automatic traffic monitoring and surveillance play an important role in the intelligent transportation system (ITS) by providing valuable information related to traffic parameters. This information would be useful both in traffic management and in assisting travellers to reach their destinations with the least time and energy consumption. At present, there are different kinds of traffic monitoring systems being deployed around the world. The traffic speed is the important parameter of traffic management which can be used to determine traffic conditions or other abnormal traffic flows. Traffic monitoring relies on traffic sensors in order to estimate these traffic parameters. There are many traffic sensors. However, video sensors offer a relatively low installation cost with little traffic disruption during installation and maintenance. They provide wide area monitoring, which implies the ability to collect a large amount of information. In addition to qualitative descriptions of traffic conditions, traffic image processing also provides quantitative traffic data such as vehicle speed, vehicle count, traffic flow, and so on. These quantitative traffic parameters are useful for traffic flow prediction and management. Other sensors, including loop detector, magnetic sensor, radar sensor, and microwave detector, suffer from serious drawbacks such as the high installation and maintenance costs as well as inability to detect slow or stationary vehicles [1]. This paper focuses on traffic speed measurement based on video sensors. In previous studies, a traffic speed measurement based on video sensors method consists of two major tasks [2]. First is object detection, the key of this task is to extract foreground object from background. There are many methods to extract foreground from background. For example, in [3] background subtraction is combined with Mixture of Gaussian (MOG) to model the precise background frame, which is used for separating moving objects from their background. Another feature-based method is space signature, where identified objects are described by their characteristics (forms, dimensions, luminosity), and this method allows identification in their environment [4]. The second task of the traffic speed measurement is object tracking, by tracking foreground object in the image sequence to describe object’s movement characteristics; i.e. trajectory and speed. Some tracking methods were proposed such as feature based tracking [4] active contour tracking [5]. But one of the main problems in this task is occlusion effect, caused by small group of vehicles or structures such as buildings or overpasses. Occlusion effect can be alleviated by Spatio- Temporal Markov Random Field (ST-MRF) model [6]. This method segments sequential images into blocks or a group of pixels. Then, an incoming vehicle is identified when intensity difference is detected, which is assigned to a new vehicle ID. The region is then updated by estimating motion vectors of the vehicle blocks. If multiple parts in vehicle region have several different motion vectors, those parts are separated and assigned to different vehicle IDs. Another method is blob tracking [7-8]. This method constructs a relation graph of blobs in current and previous frames with defined constraints. After the processes of detection and tracking are performed, velocity of each vehicle is estimated, and then the traffic speed is determined by averaging velocities from a pre- defined interval. However, there are still problems and errors in previous mentioned methods, such as false detection and false tracking due to illumination change, shadow, or complete occlusion. If these errors occur, then accuracy decreases. Moreover, the more vehicles are in region of interest (ROI), the more computation cost is required. In this paper, a new traffic speed measurement method is proposed. By projection of vehicle’s edges information in video sequence into spatio-temporal model, then this model is analysed in frequency domain using short-time Fourier transform to obtain ensemble of vehicle’s speed. This method is performed with stable computation cost even the number of