Abstract: At present there is no standard, authoritative fall detection test data, and the sample size by young people imitating fall is small, so how to use a limited data set to find the most representative feature set is particularly important. According to the characteristics of feature set in low sample and continuous type, a feature set optimization algorithm based on neighborhood consistency and discrete binary particle swarm optimization (DBPSO) was proposed. The algorithm firstly constituted the primary feature set based on optimized neighborhood consistency function and heuristic forward searching algorithm, and then used the primary feature set to initialize the population of DBPSO. At last the validity of the algorithm was verified using classification algorithm. The experimental results show that the algorithm can improve classification ability with fewer features selected, and the computational efficiency is also improved.

Key words: neighborhood consistency, DBPSO, feature selection, fall detection