Extracting relevant information regarding customer behaviour from surveillance videos

• Autor(a) principal: Correia, Simão de São José Gregório de Oliveira Frazão
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10071/29911
• topic: Customer behaviour; Object detection; Multi-object tracking; Trajectory extraction; Pose estimation; Action recognition; Comportamento do cliente; Deteção de objectos; Seguimento de objectos; Extração de trajectórias; Estimativa de pose; Reconhecimento de ações

In the modern retail environment, leveraging technologies such as high-resolution video surveillance and artificial intelligence to study in-store customer behaviour is a crucial factor in improving valuable business aspects including marketing, customer service, and security. In this context, this dissertation focuses on the development of a framework for extracting information regarding customer behaviour from high-resolution surveillance videos. This framework incorporates a series of steps, which include detecting and tracking each person, extracting trajectory points, estimating walking speeds, detecting groups, and recognising actions using pose (skeleton) data. Along with the framework, we propose two contributions designed to enhance its performance: occlusion-aware mechanism and trajectory smoothing method. The occlusion-aware mechanism was created as a means of mitigating the impact of partial occlusions on location data. The experimental results indicated statistically significant improvements resulting from its application. Furthermore, the smoothing method was introduced to attenuate oscillations in the trajectory points, considering both past and future path information. Combined with the occlusion-aware mechanism, it proved to be a valuable tool for improving trajectory mapping. Regarding action recognition, we compared three skeleton-based models (i.e. ST-GCN, AGCN, and PoseC3D) on subsets of the People in Public dataset featuring 12 shopping-related action classes. The results obtained from training and testing the models demonstrated their effectiveness in recognising customer behaviour (reaching accuracy values of around 90%) whilst ensuring privacy, and allowed us to deduce appropriate use cases for each of them.