MARINE SEARCH AND RESCUE USING LIGHT WEIGHT NEURAL NETWORKS

Abstract

The research proposes an autonomous solution using deep learning techniques inte- grated on unmanned aerial vehicles (UAV) to effect a rapid and timely search in case of a man overboard (MOB) where a person has accidentally fallen off a ship. The UAV would be deployed from a ship. A light weight neural network model is to be in- tegrated on a drone with limited processing resources. An image dataset was collected using open source videos based on people in the water (positive) and their negatives, object cropping was done by considering RGB spectrum. Three deep learning archi- tectures have been developed considering the trade-off between the minimum memory utilization (efficiency) and maximum accuracy (efficacy). It is hypothesized that a small network can achieve results equivalent to a larger network when chosen on the efficacy/efficiency trade-off. On experimental evaluation, the smaller numbers of pa- rameters yield the best F 1score performance. Neither the very large 379.7 MB ( 94 million parameters) nor very small networks 1.8 MB ( 0.03 million parameters) gave the best performance; rather the mid-size network 48.5 MB ( 12 million parameters) achieved best results.