Learning bag of visual words for motorbike detection

Recent growth of traffic surveillance based on computer vision techniques has caught more and more attentions from researchers. Since the detection of vehicles is the primary step of such system, there is a large body of works towards developing an efficient detection scheme on various operating conditions. However, those works mainly focus on the detection of car and pedestrian. In this paper, we shift our attention to motorbike, which is also a common road user, especially in developing country. In comparison to other target objects, motorbike is rather small in size but has more complex structure. Thus, detecting motorbike is not a trivial task in real-life context where there is variance due to presence of motorbike drivers and high degree of occlusions. To address this problem, we propose a method for detecting motorbike from the scenes. Our method can achieve robustness to changes in illuminations, affine transformations and occlusions. Firstly, local features are extracted from images which contain one single target object. These local features represent parts of objects and are used to construct a Bag-of-Visual Words model. Using this model, each object is represented as a histogram of their parts. Next, a Support Vector Machine classifier is trained with these representations for classifying motorbike and non-motorbike objects. Finally, we develop an algorithm to form a detection hypothesis to detect multiple target objects from the scene. We collect a dataset of 3000 images for evaluating our proposed method. The experimental results indicate that our method can achieve high accuracy in the context of real-life motorbike detection applications.