Algorithm For Early Threat Detection By Suspicious Behavior Representation

Abstract

The proposed early detection algorithm is justified because the probability of success to control a criminal activity increases when the response time for generating a warning alarm is reduced. In this paper, a video-based representation model to describe suspicious behavior from elementary actions is proposed. Such behaviors allow detecting potential threats before suspects achieve physical contact with their potential victims. In the algorithm, a novel method to adjust the balance between the anticipation level to threats and the generation of false alerts is introduced. The experimental results obtained from two validation datasets, with attacks to pedestrians and threats against a parked truck, demonstrated the effectiveness of the proposed approach for early threat detection, with performance measures above 90%.