Aggression detection in urban environments based on audio analysis

Abstract

Amid escalating security concerns, improvements in technology and reductions in costs, the Electronic Security Systems market is projected to exceed US$80 billion by 2020. Surveillance systems have become common-place within urban settings such as public transport systems, parks and motorways. Current equipment is restricted by the man power required to manually monitor and report activity, and is susceptible to human error. In this project, computer vision techniques are applied for the development of an automated aggression detection system for the detection of gunshots, breaking glass and screams in audio surveillance les. Audio surveillance sensors are a cheaper, less-intrusive alternative to conventional Closed-Circuit Television (CCTV) cameras and are able to capture events with no visual counterpart. A novel approach for the detection of aggressive events in noisy, suboptimal conditions typical of urban environments is proposed. As opposed to traditional pipelines founded on the basis of low-level audio feature analysis, this proposal has been adapted to employ computer vision techniques upon time-frequency audio representations, in order to attain a high robusticity to noise. Image features are extracted and utilized for the generation of models and the training of classi ers in order to distinguish between salient events and background noise. A stage dedicated to the localisation of events is implemented prior to classi cation, overcoming the need to segment and survey complete audio les. A comprehensive dataset composed of event recordings, superimposed with complex, realistic background noise at varying Signal-to-Noise Ratios (SNR) has been acquired for the thorough analysis of the system. The robusticity of the system to high levels of varying background noise has been con rmed, cementing the applicability of such an architecture within a forensic audio analysis context.