Efficient deep learning-based semantic mapping approach using monocular vision for resource-limited mobile robots

Potential Obstacle Detection Using RGB to Depth Image Encoder-Decoder Network: Application to Unmanned Aerial Vehicles

In this work, a new method is proposed that allows the use of a single RGB camera for the real-time detection of objects that could be potential collision sources for Unmanned Aerial Vehicles. For this purpose, a new network with an encoder-decoder architecture has been developed, which allows rapid distance estimation from a single image by performing RGB to depth mapping. Based on a comparison with other existing RGB to depth mapping methods, the proposed network achieved a satisfactory trade-off between complexity and accuracy. With only 6.3 million parameters, it achieved efficiency close to models with more than five times the number of parameters. This allows the proposed network to operate in real time. A special algorithm makes use of the distance predictions made by the network, compensating for measurement inaccuracies. The entire solution has been implemented and tested in practice in an indoor environment using a micro-drone equipped with a front-facing RGB camera. All data and source codes and pretrained network weights are available to download. Thus, one can easily reproduce the results, and the resulting solution can be tested and quickly deployed in practice.