The topic of this chapter is the implementation and analysis of vision algorithms for the detection of static and dynamic objects in videos. These algorithms are typically components of the visual perception module of modern driver-assistance systems or even autonomous cars. Whereas the vast majority of today's papers from the vision community use convolutional deep neural networks (CNNs), this chapter explores the more traditional approaches, namely HOG (Histogram of Oriented Gradients) as well as SVM (Support Vector Machine). The static and dynamic objects to be recognized are traffic signs and motorcycles, respectively. These two object classes have been chosen since traffic signs are relatively easy to detect and motorcycles state a much more complex task. Thus, this chapter tackles differently difficult tasks with a single set of algorithms.
TopIntroduction
The visual perception of the vehicle surrounding has made significant progress in the last decades. Major improvements have been achieved by introducing HOG-features (Histogram of Oriented Gradients) and SVM-classifier (Support Vector Machine) as detectors for all kinds of visual objects (see Dalal & Triggs 2005). Recent advances are the result of the introduction of convolutional neural networks, which have been presented by LeCun et al. (1998). Besides other differences, a major difference between the approaches (Dalal & Triggs, 2005) and (LeCun et al., 1998) is that the latter approach no longer utilizes a separate feature extraction step. Instead the feature extraction became in (LeCun et al., 1998) part of the neural network itself. By using a separate feature extraction step, which operates in a predefined way, i.e. it is not trained on data, this step can be optimized with respect to computational demand. Thus, the resulting detection solution will run with a higher computational efficiency. Therefore, this paper analyses the usage of (Dalal & Triggs, 2005) for the recognition of motorcycles as well as traffic signs.
The task of recognizing motorcycles is more challenging than the recognition of e.g. vehicles or lane markings, since the variation within this class is much higher. Additionally, the drivers of motorcycles can lean into a curve and therefore the recognition of motorcycles needs to take this change of position also into account.
Finding a dedicated feature descriptor for motorcycles combined with a suitable classifier is a challenging task (see e.g. Messelodi et al., 2007). The problem with the motorcycle class is increased by different motorcycle models and brands and thus various motorcycle shapes. Additionally, the driver himself increases the variation of this class. Therefore, a small moped with an overweight driver belongs to this class just as a chopper with a small driver riding it. This leads to a high-class variability. For example, Figure 1 shows a couple of examples for motorcycle models and drivers.
Using a stationary camera, it is quite easy to remove the background and use only the contour of the moving object. This is not possible with a moving camera and a moving target as this is the case for automotive applications, i.e. for driver assistance systems and autonomous driving subsystems. Another constraint for the examined approaches is
Figure 1. Sample images demonstrating the variations within the motorcycle class
the real-time capability. Especially for autonomous driving, the individual modules should run as fast as possible, in order to use the computing resources efficiently.
In the following related work to the domain of motorcycle detection will be reviewed. Many publications concerning traffic object recognition tackle vehicles, pedestrians or traffic sign recognition (see e.g. Bensrhair et al., 2001). For vehicle detection Sun et al. 2006 present a good overview of the approaches which have been used by vision researchers.
Research in the area of pedestrian detection is a very important topic for our task. This is the case because the shape of pedestrians is very similar to the shape of motorcycles. Especially it is more similar than the shape of cars, vans or trucks. Research in pedestrian detection is highly relevant for surveillance applications or for driver assistance systems. An algorithm for human body detection has been introduced by Wu et al. (2005). In their work they use socalled edgelets for classification. These edgelets define e.g. a line for the shoulder, an arc for the head-part, a symmetric pair of lines for the neck etc. Furthermore, they divide the human body into three parts (head, torso, legs) to define more suitable edgelet features. A detection of a human body is the combination of the responses of the body part detectors. In order to manage the entire edgelets in a body part, the AdaBoost algorithm was used to select the edgelets.