A SVM Based Approach for Real Time Detection and Classification of Vehicles at the Toll Gates Using Video Sequences

This paper aims to present a real-time scheme for detection and classification of vehicles passing the toll gates in Iran. In our approach, a set of videos are captured using a stationary camera, placed on the roadside, at a little distance from the toll booth. The algorithm is designed in a way that there is no need for camera calibration. Based on our videos, 3 ROIs are defined, two of them are considered to determine if a vehicle is passing and the other one is the region containing the vehicle. This work starts with the training phase, in which, for each image in a manually gathered database, HOG vectors are extracted. Two SVMs are trained in this phase, one for distinguishing vehicles from non-vehicles, and one for classifying vehicles into light and heavy vehicles. After finishing the training, in the testing phase, firstly, foreground mask is obtained differencing two consecutive frames of the video. Then, those two aforementioned ROIs are checked in every frame and as soon as a vehicle is inside the interest region, that ROI is captured. Next, the captured frame is passed to the first SVM and it is classified as vehicle or non-vehicle. Those which are identified as vehicles are passed to the second SVM to be classified as light or heavy vehicle. Average true-positive and precision rates of the vehicle detection step are 92.5% and 97.5% respectively and the same rates, for the recognition step, are 98% and 0.99%.