Desenvolvimento de um sistema inteligente para segurança rural com arquitetura flexível

Abstract

Recognizing suspicious events in videos is one of the biggest challenges in the field of computer vision. This work addresses the identification of abnormal situations in a scenario that has been little studied until then, the rural area. This scenario was selected due to the increasing crime rate in these areas and the dangers that crimes like these unleash to society. Although most of the methods proposed for recognizing abnormal events are based on supervised learning, there are some difficulties in using this type of learning in the surveillance task. Therefore, we propose the use of an event filter that receives trajectory clustering data from a genetic algorithm combined with the supervised object tracking and detection system Yolo for the recognition of abnormal situations in videos. A dataset for the chosen scenario was constructed using internet videos that proved to be good candidates for the study. Simulations of artificial trajectories were generated in order to validate the results with the real data generated. The results of this study were promising, despite the intensive processing of the genetic algorithm and the sensitivity of the tracker, the filter was able to delineate the abnormal events with both artificial and real data with satisfatory success. There is currently a great potential for boosting the results of this research into commercially viable products to help rural producers. With this study we also want to show that there is a way to shorten this path by applying the principles used in the industry within our research process. Key-words: Recognition of suspicious events on video. Unsupervised learning. Rural surveillance.