Normalized Audio Power Sequences Technique for Detecting Grain-Eating Birds Based on Bird Sounds Variance

Abstract

Monitoring events using audio sensed data has become recently an interesting research issue in multimedia based surveillance system, due to the proliferation of very tiny sensors, quite powerful to be deployed in harsh environments and to cooperatively detect critical events. Events detection can be performed through capturing, processing and decision making by an automated centralized system. Adding to this the passive nature of sound, wireless sensor networks based surveillance system may enable event detection without any requirement like line of sight, and so perform well in different application cases. In this paper, we bring out the first research element in building an automated surveillance system for rice field monitoring against pest and grain-eating birds. We study how accurate can be the use of normalized power sequences in detecting bird sounds, which is a kind of harmonic sound type. We proposed a simply detection scheme (which can be executed by sensors), for the purpose of detecting the presence of birds based on the sounds and calls they produce. The scheme is based on pruning audio frame blocks to keep relevant peaks, computing the normalized power of the sequences of captured audio files, and derive the presence of birds by only considering important values of the variances of the input files. The experimental results conducted to test the efficiency of the scheme show that 91,07% of bird calls from our database can be correctly identified.