Real-Time Active Tampering Detection of Surveillance Camera and Implementation on Digital Signal Processor

Active and automatic camera tampering detection is more and more important, especially for wide area surveillance systems with large amount of video cameras installed. Generally, a camera functions for video recording and lacks the ability to detect whether tampering has occurred. Most tampering solutions are passive, as evidenced by many surveillance systems frequently left unattended. Tampering detection of real-time automated cameras is thus of priority concern important for timely operator warning. This work describes a novel algorithm for camera tamper detection scheme. Developed for video surveillance and security applications, the pro-posed approach identifies camera tampering by detecting large edge differences and grayscale histogram comparisons between current and previous frames. The optimized algorithm is implemented on PC and TI DAVINCI DM6437 DSP as well. Extensive experiments were performed. Experimental results indicate that the proposed approach can detect various tampering cases. The overall recognition performance was 94%.

I. Introduction

Most public venues have installed many surveillance cameras, including airports, train stations, banks, schools and markets. Such surveillance systems are frequently left unattended. Tampering detection has been extensively studied. For instance, Aksay et. al. [1] introduced computationally efficient wavelet domain methods to ensure rapid detection of camera tampering and identified real-life security alarm problems. Algorithms were developed for detecting camera occlusion and reduced visibility based on a background model and wavelet transform. However background model based camera tampering detection algorithms are often unstable owing to various lighting conditions. Ribnick et. al.[2] used two stored buffer frames (short-term pool and longterm pool) as reference images to determine whether camera tampering has occurred. Nevertheless, any dissimilarity between the short-term pool and the long-term pool is flagged as tampering. Saglam and Temizel [3] provide a simple, low computational cost alternative. Pedro et. al. [4] developed an edge background model to detect camera tampering. However, for large objects or a crowd moving in front of the camera, the image characteristics are significantly altered, causing false alarms. Harasse et. al. [5] discussed a camera dysfunction detector designed mainly for surveillance systems inside vehicles. Among the other external parameters that can also influence the tampering results include illumination, weather conditions or voluntary actions. These parameters must be considered to prevent or at least detect these circumstances. This work focuses on active camera tampering detection, typically referred to as sabotage. This work presents a novel detection approach for camera tampering. An adaptive non-background model image is compared with incoming video camera frames and with an updated background image. Additionally, the moving areas of an image are monitored, along with a region based operation undertaken to reduce false alarms. This adaptive method is robust to objects moving in front of a camera. This work detects three types of camera tampering in real-time and possesses low false alarm. The proposed system is implemented on a TIDM6437 platform. Several optimization techniques, including instruction parallelism and multiple data packing in a single instruction, and floating-point arithmetic optimization, are considered. Importantly, the proposed detection algorithm is superior to conventional ones in its ability to automatically reply.