Robust vehicle detection under different weather conditions

Abstract

The introduction of new techniques for improving the robustness and accuracy of vehicle detection is always important for the Intelligent Transportation System (ITS) as they may face different problems and challenges. Variability of driving environments such as urban and interurban areas, different weather conditions, and illumination variations will affect the vehicle detection performance and later processes. This thesis proposes a number of improved methods with a high accuracy which can successfully cope with the existing challenges. In order to provide reliable results and to ensure fast and robust processing, the performance of the proposed methods have been evaluated taking advantage of an own constructed database built with videos recorded in Spain and Iran under different weather, illumination, and environmental conditions; therefore, a complete validation has been carried out. According to an exhaustive review presented in this thesis, conventional image-based vehicle detection methods have difficulties in acquiring good images due to the perspective effect and background noise. Thus, a high-accurate vehicle detection scheme by using a modified version of inverse perspective mapping (MIPM) is proposed in the first method. A cooperative framework is adapted based on two complementary analyses. The spectral features which follow a spatial feature are integrated for the pixel-level analysis. Subsequently, the regional-level analysis takes place to extract the geometric features of vehicles and reducing the false-positive rate. The later analysis is performed at the original image based on a discrete differentiation operator in order to get accurate information of vehicles. Thereupon, rule-based strategies are proposed to reduced the false-negative rate and provide the final detection results. As experimental results indicated, higher accuracy is achieved by means of MIPM transformed image leading to increase vehicle detection performance under all weather conditions. Experimentally, we verified that small qualitative differences of transformation methods can manifest themselves crucially important in traffic monitoring systems. Comparing with previously published methods it demonstrates that MIPM, was not only simple and straightforward, but it was also more accurate in front of others. On the other hand, the proposed method opens the door to extract and collect real traffic information. In order to deal with a camera vibration due to the wind or bridge movements, adverse weather conditions, and achieve an accurate vehicle detection system, the second method is proposed. This has been developed through three main units which are tightly integrated. Particularly, the motion-based unit demonstrates the existence of vehicles by means of spectral and temporal features based on statistical characteristics of image pixels and a combination of several extra processing. Moreover, the accuracy of the system will be increased by the effective incorporation of the acquired data at the different parts of the proposed method. Regarding the knowledge-based unit, an adaptive particle filter framework with multiple measurements and a realistic noise model was developed which requires the information from the prior unit at the different stages. As a consequence, the second unit is beneficial as it allows refining the coarse results obtained by the former which happened due to the unfavorable conditions. Finally, the decision-making unit increased the accuracy of inter-frame correspondence by considering the various situations from two different points of view. The evaluation is divided into two extensive parts to prove the higher performance of the proposed method in front of others. Different from previous literature, the proposed method achieved a more accurate result under bad weather conditions by the effective integration of diverse information. In summary, experimental results show that the proposed methods offer great improvements in terms of accuracy, robustness, and stability in traffic surveillance. KEYWORDS: ITS, MIPM, Vehicle detection, Detection accuracy, Particle filter, Vibrating camera, Adverse weather conditions.