This thesis presents my contributions to the state-of-the-art in Intelligent Transportation Systems and Computer Vision. Specifically, the work is focused on two tasks -- automatic speed measurement of vehicles and fine-grained recognition of vehicles. 

The problem of vehicle fine-grained recognition can be defined as a task where the system is supposed to produce exact fine-grained type (e.g. "Škoda Octavia combi mk2") for a presented vehicle. In my first paper on this topic, a method exploiting automatically constructed 3D bounding boxes around vehicles is proposed. The results show that the method significantly improves classification and verification accuracy. Further improvements and analysis of the approach was published in my second paper dealing with the problem. The improved approach eliminates necessity to know vanishing points a priori - it is possible to construct the 3D bounding box of the vehicle from a single image of the vehicle. The results show that our proposed method consistently improves classification accuracy by up to 12 percentage points with different Convolutional Neural Networks. The classification error was also reduced by up to 50 %.The second addressed problem is automatic speed measurement of vehicles. The proposed system should work from a single camera without any manual calibration or input. First, we had to collect a large dataset with precise ground truth speed measurements as there was no such dataset. The dataset contains over 20,000 vehicles with ground truth speed measurement acquired from two synchronized LIDAR optical gates. Furthermore, we proposed a method for fully automatic traffic surveillance camera calibration enabling precise speed measurement of vehicles. The approach is based on vanishing point estimation and 3D model alignment of several common vehicle models. The experimental results show that our method achieves 1.10 km/h mean speed measurement error while outperforming both state-of-the-art methods and manual calibration in the speed measurement task.