Image-based Obstacle Avoidance using 3DConv Network for Rocky Environment

Autonomous navigation systems are an essential part of Unmanned Ground Vehicles (UGVs) since they allow for navigating without supervision in conditions where communication is not available or the existence of high delays which prevent direct communication. One fundamental part of autonomous navigation is obstacle avoidance. Typical approaches utilize some form of distance measuring-based sensors like LIDAR or SONAR. However, such devices have a relatively higher cost in comparison to conventional RGB cameras in addition to introducing complexity in data processing which results in an increase in computational cost and power consumption. In this work, we use sequential RGB data and a Conv3d-based network to create a real-time obstacle avoidance system with high accuracy, low latency, and low processing cost. For training, we used Unreal Engine 4 based simulator to collect a dataset to train the network. Testing the system in a simulated environment using the same simulator showed the ability of the network to avoid obstacles in a realistic environment where rocks of different sizes and shapes were used. Future work can include improving in terms of performance and processing time as well as implementing the network with a real word working prototype and comparing the simulated results with actual performance.