Urban traffic is one of the most important causes of noise in the cities. Noise metrics, such as Equivalent and Day-night levels, are based on road traffic characteristics over long term time periods and do not describe the traffic noise dynamics. This paper presents a method for predicting and producing dynamic noise maps from flow traffic data obtained by surveillance cameras and processed by machine learning algorithms. Convolutional networks, based on transfer learning process, are used to identify the vehicles in the scene and classify them according to their size, while geometry projection algorithms are used to determine the vehicle position, speed and acceleration. The data extracted from video, such as vehicle category, speed and lane, are useful to calculate sound pressure level in any given location. Therefore, by adding energetically such levels for each frame, a dynamic noise map can by produced, representing the sound pressure variation over the time, according to the traffic flow. The levels predicted from the videos are compared to measurements realized in control points to validate the model. Results of the proposed model are presented as well as the model restrictions to produce the noise maps.