Abstract
Building a fight detection system is essential to enhance public safety and develop automated security surveillance systems. Given the success of deep learning-based models in computer vision and pattern recognition, it is common to use deep convolutional and recurrent neural networks to accurately predict fight scenes. However, the required computational resources hinder the use of these models in memory-limited devices. In this work, we aim to obtain the fight detection model that satisfies both memory constraints and accurate classification performance. First, we design a complex model and then reduce its complexity by leveraging a three-stage progressive model optimization strategy. As a result, we obtained accurate detection models, which are $\sim 2$ and $\sim 5$ times lighter compared to the starting complex model. The experiments are performed using two publicly available datasets, namely Hockey and Movie Fights.