RGB-D Scene Understanding And Its Optimization

RGB-D scene understanding is extremely valuable to autonomous car and indoor robotnavigation. Accurate depths information provided by RGB-D sensors can compensate theweaknesses of stereo methods, and let researchers focus on high level problems, such asobject recognition and scene classification.With sparse connectivity and weight sharing, convolutional neural networks are verysuitable for solving complex vision problems, such as3D scene understanding. Training aconvolutional neural network is usually transformed into a single objective optimizationproblem, but neural network training is multiobjective by nature.This thesis proposes a new multiobjective particle swarm algorithm, which is able tomaintain a set of infeasible solutions to accelerate the convergence rate. A method forcalculating fitness values for infeasible solutions is designed. The fitness of an infeasiblesolution is affected by the amount of constraint violation, density in the objective space andfeedback from the particle update process. Density is estimated by an improved adaptive grid,while feedback is achieved by a voting mechanism.The training of convolutional neural networks is treated as a multiobjective optimizationproblem, and solved by the new multiobjective particle swarm algorithm. Then, combinedwith a new normal vector calculation method, the multiobjective convolutional neuralnetwork is applied to RGB-D scene understanding problems, and its application inautonomous car and indoor robot navigation is discussed.Experiment results show that the new multiobjective particle swarm algorithm iscompetitive, and that multiobjective convolutional neural network based RGB-D sceneunderstanding is very valuable to applications.