Automatic video surveillance systems have been developed to detect and analyze abnormal behavior or situation of risk in many fields reducing human monitoring of activities captured by cameras (security surveillance, abnormal behavior detection, etc.). One of the applications of video surveillance is the traffic monitoring. Analyzing the motion in roads aims to detect abnormal traffic behavior and sudden events, especially in case of Emergency and Disaster Management (EDM). Road accidents can cause serious injuries affecting mostly the head and the brain, leading to lifelong disabilities and even death; each additional rescue minute can mean the difference between life and death as revealed by the golden Hour[Lerner et al., 2001]. Therefore, providing a rapid assistance for injuries is mandatory. Moreover, if not addressed promptly, accidents may cause traffic jams, eventually leading to more accidents, and even greater loss of lives and properties. Many cities in France are equipped with video surveillance cameras installed on different roads and highways. Traffic monitoring is done by human operators to visualize the congestion of a road or to measure the flow of the traffic. The video stream of this existing network of cameras is delivered unprocessed to the traffic management center. Thus, there are no video storage of accident scenes. In addition, there is no associated technology for a rapid emergency management. Therefore, it is important to design a system able toorganizean effective emergency response. This response should be based, firstly on an automatic detection by video analysis, then, on a rapid notification allowing the optimization of the emergency intervention itinerary without affecting the traffic state. Our work resolves the first part of the emergency response.The objectives of this thesis are firstly the identification of accident scenarios and the collection of data related to road accident; next, the design and the development of video processing algorithms for the automatic detection of accidents in highways. The developed solutions will use the existing fixed cameras, so as not to require significant investments in infrastructure. The core of the proposed approaches will focus on the use of the dense Optical Flow (OF) algorithm [Farnebäck, 2003] and heuristic computations for features extraction and accident recognition. The purpose of the dense OF is to estimate the motion of each pixel in a region of interest (ROI) between two given frames. At the output of the dense OF, a dense features could be extracted which is more performant than features extracted at some points. Defining thresholds for accident detection in various environment is very challenging. Therefore, studying the motion at a global scale in the image, allows defining a dynamic thresholds for accident detection using statistic computations. The proposed solution is sufficient and robust to noise and light changing.