The indoor sweeping robot industry has developed rapidly in recent years. The current sweeping robot environment perception sensor configuration is more diverse and generally does not have active garbage detection capabilities. Advances in computer vision technology, artificial intelligence, and cloud computing technology have provided new possibilities for the development of sweeping robot technology. This paper conceptualizes a new autonomous visual navigation system based on a single-camera sensor for floor-sweeping robots. It investigates critical technologies such as floor litter recognition, environmental perception, and dynamic local path planning based on depth maps. The system is applied to the TurtleBot robot for experiments, and the results show that the mAP accuracy of the autonomous visual navigation system for fine trash recognition is 91.28%; it reduces the average relative error of depth perception by 10.4% compared to conventional methods. Moreover, it has greatly improved the dynamics and immediacy of path planning.