Spatio-Temporal Vehicle Tracking Using Unsupervised Learning-Based Segmentation and Object Tracking

Introduction Recently, Intelligent Transportation Systems (ITS), which among others make use of advanced sensor systems for on-line surveillance to gather detailed information on traffic conditions, have been identified as the new paradigm to address the growing mobility problems. With the exponential growth in computational capability and information technology, traffic monitoring and large-scale data collection have been enabled through the use of new sensor technologies. One ITS technology, Advanced Traffic Management Systems (ATMS) [1], aims at using advanced sensor systems for on-line surveillance and detailed information gathering on traffic conditions. Robotic vision [2], especially 2D imaging based vision (2D image processing, object tracking, etc.), can be applied to traffic video analysis to address queue detection, vehicle classification, and vehicle counting. In particular, vehicle classification and vehicle tracking have been extensively investigated [3][4]. Issues associated with extracting traffic movement and accident information from real-time video sequences are discussed in [5]. For traffic intersection monitoring, digital cameras are fixed and installed above the area of the intersection. A classic technique to identify the moving objects (vehicles) is background subtraction [6]. Various approaches to background subtraction and modeling techniques have been discussed in the literature [4][5]. In the proposed framework, an unsupervised video segmentation method called the Simultaneous Partition and Class Parameter Estimation (SPCPE) algorithm is applied to identify the vehicle objects in the video sequence [7]. In addition, we propose a new method for background learning and subtraction to enhance the basic SPCPE algorithm in order to generate more accurate segmentation results, so that more accurate spatio-temporal relationships of objects can be obtained. Experiments are conducted using real-life traffic video sequences from road intersections. The experimental results indicate that almost all moving vehicle objects can be successfully identified at a very early stage of the processing, thereby ensuring that accurate spatio-temporal information of objects can be obtained through object tracking.