Visual Surveying of On-Road Vehicle Height for Over-Height Warning Using Deep Learning and View Geometry

Computer vision-based over-height vehicle detection has received increasing attention for preventing vehicle-to-structure collisions. However, the prior studies usually rely on blob detection of vehicles for vehicle height measurement, which is oftentimes error-prone because the traditional blob detection methods are sensitive to shadow interference and illumination change. Also, it is practically difficult to extract the vehicle’s pixel height from the detected blobs automatically and accurately due to the lack of depth information on 2D images, especially for vehicles with irregular top shapes. Furthermore, the applications of the existing methods are still plagued by the occlusion issue (e.g., part of a vehicle is blocked by another vehicle) that commonly occurred in surveilled views. Therefore, this research proposed a new method for vehicle height estimation. In the proposed method, the vehicle instances are first segmented from images using Mask R-CNN, which can significantly eliminate the adverse effect of shadow interference and illumination change. Then, 3D bounding box is established for every surveilled vehicle with or without being partially blocked by leveraging the identified vehicle instance and the three orthogonal vanishing points in the surveilled scene. By doing so, the vertical edges of established 3D bounding box are directly associated with the vehicle’s height, thus overcoming the difficulty of identifying the vehicle’s pixel height in 2D images. Finally, the vehicle’s physical height is estimated using a known object length in the same surveilled scene according to view geometry. The evaluation results signify the potential of the proposed method for use in over-height collision prevention.