Estimation of moving agents density in 2D space based on LSTM neural network

Abstract

As a part of ASSISIbf project, with a final goal of forming a collective adaptive bio-hybrid society of animals and robots, an artificial neural network based on LSTM architecture was designed and trained for bee density estimation. During experiments, the bees are placed inside a plastic arena covered with wax, where they interact with and adapt to specialized static robotic units, CASUs, designed specially for this project. In order to interact with honeybees, the CASUs require the capability i) to produce and perceive the stimuli, i.e., environmental cues, that are relevant to honeybee behaviour, and ii) to sense the honeybees presence. The second requirement is implemented through 6 proximity sensors mounted on the upper part of CASU. In this paper we present estimation of honeybees (moving agents) density in 2D space (experimental arena) that is based on LSTM neural network. When compared to previous work done in this field, experiments demonstrate satisfactory results in estimating sizes of bee groups placed in the arena within a larger scope of outputs. Two different approaches were tested: regression and classification, with classification yielding higher accuracy.