Deep speed estimation from synthetic and monocular data

Current state-of-the-art in speed measurement technologies includes magnetic inductive loop detectors, Doppler radar, infrared sensors, and laser sensors. Many of these systems rely on intrusive methods that require intricate installation and maintenance processes that hinder traffic while leading to high acquisition and maintenance costs. Speed measurement from monocular videos appears as an alternative in this context. However, most of these systems present as a drawback the requirement of camera calibration - a fundamental step to convert the vehicle speed from pixels per frame to some real-world unit of measurement (e.g. km/h). Considering that, we propose a speed measurement system based on monocular cameras with no need for calibration. Our proposed system was trained from a synthetic data set containing 12,290 instances of vehicle speeds. We extract the motion information of the vehicles that pass in a specific region of the image by using dense optical flow, using it as input to a regressor based on a customized VGG-16 network. The performance of our method was evaluated over the Luvizon's data set, which contains real-world scenarios with 7,766 vehicle speeds, ground-truthed by a high precision system based on properly calibrated and approved inductive loop detectors. Our proposed system was able to measure 85.4% of the speed instances within an error range of [-3, + 2] km/h, which is ideally defined by the regulatory authorities in several countries. Our proposed system does not rely on any distance measurements in the real world as input, eliminating the need for camera calibration.