Monocular Visual Odometry Based on Deep Convolutional Networks with Image Enhancement and Attention Mechanisms

In recent years, the research focus of unmanned vehicle positioning has gradually shifted to using visual information to estimate camera pose, in which visual odometer plays a key role. Under the background of deep learning, on the one hand, monocular vision odometer method shows potential. In order to solve the problem of insufficient features in low texture environment, traditional geometric algorithm is embedded in deep learn-based VO to enhance the influence of edge feature information in images, an image preprocessing method combining edge feature extraction is proposed. On the other hand, the channel domain attention mechanism is integrated into the feature extraction stage, and then the timing modeling is completed using the short-short network to achieve the end-to-end output pose. Experimental results show that this method performs well in low texture environment.