The Comparison of Light Compensation Method and CycleGAN Method for Deep Learning Based Object Detection of Mobile Robot Vision Under Inconsistent Illumination Conditions in Virtual Environment

The visual system of indoor mobile robots is mainly used for object detection, recognition, tracking, positioning and navigation. However, the effect of object detection of robot is easily affected by the outside world such as: illumination inconsistent, object occlusion, and so on. In order to explore and solve the effect of illumination inconsistent on object detection algorithm(YOLOv2) of robot vision, the virtual environment was used to simulate the real indoor environment, because the conditions are not easily controlled and susceptible to external influences for object detection in the actual indoor environment. And the low confidence of object detection of robot vision was studied and verified in the dark environment. Meanwhile, the light source was added to the robot to solve relevant problems and specific method of brightness migration based on CycleGAN is found after analyzing the principle. Finally, the method based on CycleGAN is applied to robot to improve the detection effect in dark environment.