The 3D capability map model of humanoid robot based on 3D point cloud data has the disadvantage of large voxel mesh searching computation. Considering the hierarchical advantage of OcTree in 3D space subdivision, a local environment and capability map model based on Octomap was proposed. Firstly, a binary-tree-like kinematics model of NAO humanoid robot was constructed according to the joint composition, forward kinematics, inverse kinematics and rigid body coordinate transformation of NAO robot. Secondly, the forward kinematics was used to calculate the 3D discrete reachable point clouds in Cartesian space, which were used as the basic workspace of the robot terminal effector. Thirdly, the methods of transforming the point cloud space representation into Octomap space node representation, especially the probability updating method of space node, were described emphatically. Finally, an optimization method of space node updating order selection was proposed according to the geometric relationship of nodes. With this optimization method, the space optimization representation of the humanoid robot's capability map was realized efficiently. Experimental results show that compared with the original Octomap updating method, the proposed algorithm can reduce the number of space nodes by nearly 30% and improve the computional efficiency.

Aiming at the problem that biped robots may easily lose their motion stability when walking on uneven ground, a value-based deep reinforcement learning algorithm called Deep Q-Network (DQN) gait control method was proposed, which is an intelligent learning method of posture adjustment. Firstly, an off-line gait for a flat ground environment was obtained through the gait planning of the robot. Secondly, instead of implementing a complex dynamic model compared to traditional control methods, a bipedal robot was regarded as an agent to establish robot environment space, state space, action space and Reward-Punishment (RP) mechanism. Finally, through multiple rounds of training, the biped robot learned to adjust its posture on the uneven ground and ensures the stability of walking. The performance and effectiveness of the proposed algorithm was validated in a V-Rep simulation environment. The results demonstrate that the biped robot's lateral tile angle is less than 3° after implementing the proposed method and the walking stability is improved obviously, which achieves the robot's posture adjustment behavior learning and proves the effectiveness of the method.

To address the underperformance of feature-based image registration under situations with large affine deformation and similar targets, and reduce the time cost, an improved Descriptor-Nets (D-Nets) algorithm based on matching quality purification was proposed. The feature points were detected by Features From Accelerated Segment Test (FAST) algorithm initially, and then they were filtered according to Harris corner response function and meshing. Furthermore, on the basis of calculating the line-descriptor, a hash table and a vote were constructed, thus rough-matching pairs could be obtained. Eventually, mismatches were eliminated by the purification based on matching quality. Experiments were carried out on Mikolajczyk standard image data set of Oxford University. Results show that the proposed improved D-Nets algorithm has an average registration accuracy of 92.2% and an average time cost of 2.48 s under large variation of scale, parallax and light. Compared to Scale-Invariant Feature Transform (SIFT), Affine-SIFT (ASIFT), original D-Nets algorithms, the improved algorithm has a similar registration accuracy with the original algorithm but with up to 80 times speed boost, and it has the best robustness which significantly outperforms SIFT and ASIFT, which is practical for image registration applications.