Monocular Reconstruction of Vehicles: Combining SLAM with Shape Priors

1 minute read

Abstract

Current approaches leverage two kinds of information to deal with the
vehicle detection and tracking problem
- (1) 3D representations (eg. wireframe models or voxel based or CAD models) for diverse vehicle skeletal structures learnt from data,and
- (2) classifiers trained to detect vehicles or vehicle parts in single images
  built on top of a basic feature extraction step
First, we extend detection to a multiple view setting
Secondly, we can also leverage 3D information from the scene generated using a unique structure from motion algorithm.

In the present paper, we tightly integrate these deformable 3D object models with state-of-the-art multibody SfM methods,
Approximating the visible surfaces of a vehicle by planar segments, supported
by discriminative part detectors allows us to obtain more stable and
accurate 3D reconstruction of moving objects as compared to state-of-the-art SLAM pipelines
Summarily, we list the contributions of the present paper in the following

We propose a novel piece-wise planar approximation to vehicle surfaces and use it for robust camera trajectory estimation. The object presents itself as a plane to the moving camera. By segmenting the car into its constituent planes by RANSAC with Homography as the model we obtain superior reconstruction of the moving object
We extend the single-view deformable wireframe model fitting to multiple views, which stabilizes the estimation of object location and shape
We experimentally demonstrate improvements in 3D shape estimation and localization on several sequences in KITTI dataset [13] resulting from the the tight integration between SfM cues and object shape modeling