Emprego de visão computacional para identificação de comportamento suspeito no varejo

Abstract

This work addresses the chronic problem of inventory losses due to theft in the retail sector, aiming to empirically compare the performance of two known deep learning architectures for the automated detection of suspicious behavior in videos: a Three-Dimensional Con- volutional Neural Network (I3D) and a Vision Transformer (TimeSFormer). To conduct the study, the methodology involved building a proprietary dataset through the aggrega- tion, filtering, and segmentation of samples from public datasets, subjecting the models to sixteen experimental configurations. These experiments varied the weight fine-tuning strategy, temporal resolution, and the use of optical flow images combined with standard RGB inputs. The evaluation also included a simulation of continuous monitoring on real, unedited videos using a sliding window inference approach. The results demonstrate that full fine-tuning of all network parameters is a strictly essential step for convergence and learning in both architectures. The Vision Transformer achieved the best generalization capability and false alarm control on the test set, attaining perfect precision; however, it requires substantially higher computational investment and exhibits high inference time. In contrast, the Convolutional Neural Network showed a more conservative predictive behavior with a greater tendency toward false positives, while standing out for its high efficiency and low latency. It is concluded that practical deployment requires a direct tech- nical trade-off between the high selectivity of attention-based models and the scalability and speed of convolutional networks. Additionally, both systems share critical vulnera- bilities and detection failures when faced with severe occlusions and subtle movements in retail environments.