Multi-Objective Deep CNN for Outdoor Auto-Navigation

Target-guided navigation establishes the foundation for efficiently addressing vision-based multi-agent coordination for robotics. This work proposes a multi-objective deep convolution network which consists of two parallel branches built atop a shared feature extractor. The proposed network is capable of concurrently constructing semantic maps while achieving efficient visual detection of a designated guider robot or landmark towards outdoor navigation. In order to achieve the low latency requirements of the navigation controller, the structure and parameters of the network have been meticulously designed to boost run-time performance. The model is trained and tested on an altered version of the Cityscape outdoor dataset. We further finetune using a collected dataset in order to improve generalization performance on unseen outdoor scenes. Experimental results on an outdoor navigation robot equipped with an RGBD camera and GPU mini PC verifies the feasibility of the model.