Deep Direct Visual Odometry

Traditional monocular direct visual odometry (DVO) is one of the most famous methods to estimate ego-motion robots and map environments from images simultaneously. However, DVO heavily relies on high-quality accurate initial pose estimation during tracking. With outstanding performance deep learning, previous works have shown that neural networks can effectively learn 6-DoF (Degree Freedom) poses between frames image sequences in unsupervised manner. these learning-based frameworks cannot accurately generate full trajectory a long video because scale-inconsistency each pose. To address this problem, we use several geometric constraints improve scale-consistency network, including improving loss function proposing novel scale-to-trajectory constraint for training. We call network trained by proposed as TrajNet. In addition, new architecture, called sparse (DDSO), overcome drawbacks (DSO) framework embedding Extensive experiments KITTI dataset show TrajNet when compared with methods, integration makes initialization tracking DSO more robust accurate.