Multi-UAV Maritime Search and Rescue with DNN Inference Acceleration

Unmanned aerial vehicle (UAVs) bring great advantages to maritime search and rescue (SAR), but UAVs with limited computing power and memory resources cannot meet the strict deadline requirements when processing computationally intensive deep neural network (DNN) tasks. In order to overcome the related resource constraints, DNN inference tasks are generally offloaded to the cloud or nearby powerful nodes. However, the existing work fails to effectively arrange resources for DNN inference tasks in a cluster environment. Here, we innovatively proposed DNN inference acceleration (DIA), an approach based on DNN fine-grained partitioning and offloading, to accelerate DNN inference in cluster environment. Specifically, we study the parallel execution of DNN inference on multiple devices. The proposed strategy includes a fine-grained partitioning algorithm of DNN (DIA-FP), and an offloading algorithm (DIA-O) based on the overlapping coalition structure formation method. Results obtained by using a large-scale dataset of ships and four DNN benchmarks show that compared with other partitioning and offloading methods, the proposed method significantly reduces the total inference time of DNN, and the inference speed is
$\mathbf{1}.\mathbf{2}\times$
to
$\mathbf{2}.\mathbf{9}\times$
faster.