Blur Detection for Surveillance Camera System

Surveillance cameras, which are often placed in unconstrained environments, can be tampered with due to many environmental and human factors. It results in degraded surveillance videos and affects the subsequent smart applications that make use of the videos in decision-making. Blur anomaly is one of the most typical problems in those videos, which have the target objects in the videos significantly blurred or partially occluded. Such blur anomalies should be detected as soon as possible to ensure the integrity of the video data. In this study, a novel self-supervised blur detection model is proposed. The model focuses on the detection of four commonly found blur anomalies in surveillance videos. They are the natural blur, defocus blur, dirt blur, and spray-paint blur. By using the self-supervised learning method, we can fully make use of the abundant positive samples to improve detection accuracy. Since, in many situations, camera anomaly detection needs to be carried out with edge devices, we also propose a compressed residual network, which incorporates a color attention module, to reduce the model complexity so that it can be applied to edge applications. Our experimental results show that the proposed model significantly improves over the existing approaches in terms of complexity and accuracy.

I. Introduction

Surveillance camera systems are commonly used nowadays for security monitoring, activity detection, and object detection [1]–[11]. These systems can have thousands of cameras, and they are often installed in places that are difficult to access. It is hard to continue monitoring their working conditions manually. Surveillance cameras can degrade due to many reasons, including natural influence and manual sabotage. Since most of these cameras are mounted outdoors, the captured videos can be affected by extreme weather conditions such as heavy rain, snow, fog, dust, etc. The quality of the captured video will also degrade due to the aging of the camera itself or the accumulated dirt on the camera lens. Besides, human intruders may spray-paint the camera lens with an aim to reduce the visibility of the cameras. All the abovementioned factors will generate blurring or partial occlusion effects on the surveillance images. We call them blur anomalies. When using these degraded surveillance images in smart applications, wrong decisions will be made. Therefore, automatic detection of these degraded surveillance images is required to ensure the integrity of the surveillance video data. In this study, we focus on the following four types of blur anomalies: natural blur, defocus blur, dirt blur, and spray-paint blur. The natural blur refers to the blurring effect caused by environmental changes, such as heavy rain, snow, fog, dust, etc. The defocus blur can be generated due to the aging of the camera. The dirt blur is introduced by the dirt on the camera lens. The spray paint blur is found when intruders spray paint onto a surveillance camera with an intention to block its view Some examples of the blur anomalies are shown in Fig. 1. Our target is to develop efficient learning-based methods to detect these blur anomalies. Since a surveillance camera system can have thousands of cameras, the complexity of the blur detection method needs to be well controlled so as not to overload the system server. It may even be better if the blur detection can be carried out on the edges so as to remove the burden of the system server for doing other classification tasks. Thus, low complexity blur detection is the objective of this study.