Precision-enhanced object recognition for semantic SLAM fusion

Abstract

Recent developments in deep learning techniques have accelerated the growth of robotic vision systems for many applications. This thesis examines several object detection algorithms that can be combined with simultaneous localization and mapping (SLAM) to produce an annotated 3D map of the local environment. After the initial review, Deep Object Pose Estimation (DOPE) is examined further and modified to enable object recognition using a depth map as input. This modification allows for shape-based object recognition in low visibility conditions such as darkness or camouflage. The method starts with a dual threshold to isolate a detection window followed by median and sharpness filtering to reduce the noise in the depth map. Moreover, a novel method to quantify the uncertainty of predictions is used to evaluate a threshold applied to key points used to enhance precision of the algorithm. Precision of the method surpasses the baseline across all frame rates tested. The method maintains high precision when the depth of the convolution neural network is reduced from 86 layers to 21 layers promoting an increase in frame rate from 8.6 Hz to 22.6 Hz. Additionally, five datasets are tested to determine the optimal size that should be used for training. Results show that 100k images is ideal, achieving the highest position accuracy of 96.8%, and that overfitting becomes apparent above 300k images as accuracy starts to drop. Finally, DOPE is combined with Real-Time Appearance-Based Mapping (RTAB-Map) and the enhanced precision method is compared against the original version to evaluate object placement on the map generated by SLAM. The original version yields an average error of 79.8 cm distance from the ground truth compared to 30.5 cm using the proposed method.