Classification of Vehicles using Monocular 3D Reconstruction

State of the art 3D reconstruction techniques utilize frames from a video sequence to render a 3D model of the scene. Our 3D reconstruction technique utilizes Speeded-Up Robust Features along with optical flow points to create a dense point cloud. Each point within the model has been tracked from frame to frame and triangulated into its (X,Y,Z) model position. We present an application for these structure from motion models that exploits our previous work in 3D object classification. In our experiments, we reconstruct a parking lot scene that contains several vehicles. The first step of our object classification algorithm is to segment each of the vehicles. Then, for each separate point cluster, our algorithm utilizes the volumetric and shape properties of the 3D object to label it with a vehicle type. The novelty of this classification approach allows us to tackle the noise challenges commonly associated with monocular 3D reconstructed models.https://ecommons.udayton.edu/stander_posters/1652/thumbnail.jp