occluding vehicles may be grouped together as one object under heavy congested traffic conditions. A Hidden Markov Model is used in [7] to detect cars under congested conditions, where occlusion is frequent. A different approach, presented in [13], does not require the background estimation, but instead it detects information about the corners of the vehicles. Both [9] and [17] present a good survey of other vehicle detection algorithms, in which the latter proposes a way to combine several of these methods.Another common task is shadow removal, with several benefits. In fact, the shadow of a moving object is also moving and hence could be considered a moving object. If the shadow of a car is considered together with the car, it may seem that the car is bigger than it really is. Another situation to bear in mind is when the shadow of a car connects to other cars, causing not optimized algorithm to merge both cars into one single object. A solution to remove shadows from images is then presented in [10], which uses top-hat transformations and morphological operators. In [6], a solution to remove shadows based on a single Gaussian shadow model is presented, whereas [18] describes other alternatives with references and further information about them.Estimating the speed of vehicles is an important task be- cause it can be used as a source of information for most of the traffic monitoring applications. This task is challenging owing to the difficulty in acquiring, lack of precision of means used and highly sensible data. Not even humans, observing attentively videos with good quality, can precisely tell which pixels belong to one car or another. Also, blur from cars moving at high speeds makes this even more difficult. Indeed, the exact position of the camera is not known (and cannot be calculated very precisely) besides vibrations and movement of the camera that cause unpredictable variations. All these difficulties together cause enormous noise when geometrical information that relies on the position of the camera and observed cars is required. This happens because the applied methods are highly sensible. In other words, small changes in one variable may induce into big differences in the final estimated velocity. In [9], authors estimate speed based on the point of contact of a car with the road, whereas [10] uses a geometrical equation to accomplish that.Tracking is also one important application, being vehicle tracking the most common type. Nonetheless, it is also pos- sible to track other features of an image. For a comprehen- sive discussion on tracking algorithms, the interested reader is referred to [12], which also suggests a methodology to track vehicles using a scale invariant feature transform. This algorithm, applied to each vehicle, will describe its features such as pixel values, key point locations and orientations into a 128 dimensional vector, which can be tracked in the following frames. In [13], instead, the suggested technique detects corners of vehicles, tracks them using Kalman filtering and groups the corners into vehicles.