Micro-UAV onboard vehicle detection: architecture and experiments

Abstract

In this paper, an onboard processing architecture is proposed for micro fixed-wing unmanned aerial vehicle (UAV). The typical application scenarios include online detection of moving vehicles on the ground. The detection architecture is compatible with extremely limited computing resources provided by the micro UAVs. Eventually, the multi-vehicle detector is composed of saliency-based region proposal and neural network supported classifier. A typical convolutional neural network, connected with the Cifar10 dataset [1], is selected as the classifier, by performance comparisons driven by the annotated datasets. Furthermore, both “progress balance” and “quantity balance” strategies are developed and compared for training sample structure optimizing to reduce misclassification and leakage classification. Under such circumstances, experiments are conducted with onboard imagery of flying micro fixed-wing vehicles during surveillance on multiple moving vehicles on the ground. Experimental results validate the feasibility and effectiveness of the proposed onboard detection architecture. Typically, four-vehicle mAP is promoted from 52.50% to 57.76% by using the unified progress and quantity dataset balance strategy.