Counting pedestrians and bicycles in traffic scenes

Object detection and classification have received increased attention recently from computer vision and image processing researchers. Image processing views this problem at a much lower level as compared to machine learning and linear algebraic analysis which focus on the overall statistics of object classes given sufficient data. A good algorithm uses both these approaches to its advantage. It is important to define and choose the features of an image suitably, so that the classification algorithm can perform at its best in distinguishing object classes. In this paper we investigate the performance of different types of texture-based features when used with a support vector machine. Their performance was evaluated on standardized image datasets and compared. The objective of this study was to come up with a suitable algorithm to distinguish bicycles from pedestrians in locations such as bicycle paths and trails in order to estimate their traffic. The models developed during this study were applied in practice to traffic videos and the results are summarized here. For better application in practice other cues derived from motion were utilized to improve the performance of the classification and hence the accuracy of the counts.