基于SOLOv2-RS的人工假体视觉避障研究

Obstacle avoidance in simulated prosthetic vision based on SOLOv2-RS

面向人工假体视觉条件下的避障问题,提出改进的实例分割模型SOLOv2-RS,为植入者在低分辨率人工视觉中更准确地感知导航任务的相关实例对象提供基础。根据视觉注意力机制,采取视野中心距离和目标尺度作为各实例重要性计算准则,以得到的重要性分数作为对需规避障碍物进行分级表达的依据;同时,采用边缘信息提示盲道,并对其进行形态学膨胀处理以避免光幻视有限导致的边缘信息缺失。人工假体视觉仿真结果表明本研究提出的人工假体视觉分级优化处理策略能有效实现盲道和障碍物的优化表达,为植入者更高效地完成室外避障任务提供便利,为人工假体视觉设备图像处理研究提供良好思路。

Aiming at the obstacle avoidance in simulated prosthetic vision,an improved instance segmentation model SOLOv2-RS is proposed for providing a basis for implant recipients to accurately perceive the relevant instance objects of navigation tasks in low-resolution prosthetic vision.According to the visual attention mechanism,the distance from the center of the visual field and the target scale are adopted as the importance calculation criteria for each instance,and the obtained importance score is used as the basis for the hierarchical representation of the obstacles to be avoided.Meanwhile,edge information is used to cue the tactile paving,and it is morphologically inflated for avoiding the edge information loss caused by the limited phosphene.The prosthetic vision simulation results demonstrate that the hierarchical optimization processing strategy for simulated prosthetic vision can effectively achieve the optimal representation of tactile paving and obstacles,thus facilitating the implant recipients to accomplish outdoor obstacle avoidance tasks more efficiently,and providing ideas for the research on the image processing of visual prosthetic devices.