1.1. Evolution of Sensors & Analytics Engine

Over the last four decades, closed-circuit television (CCTV) systems have evolved from the first to the current third generation, incorporating intelligent features [1], [2], [3]. Progress in the system's analytics engines, classified into Computer Vision (CV)-based [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16] and Artificial Intelligence (AI)-based [17], [18], [19], [20], [21], [22], [23], [24] categories, primarily contributes to these advancements.

Classical CV-based surveillance systems continue to be prevalent despite the rise of AI. These systems utilize optical imaging sensors and spatial feature-based processing for anomaly detection. To illustrate, Tiwari and Verma applied optical sensors, color-based segmentation, and the Harris interest point detector for weapon detection [4]. Sanoob et al. [5] used optical imaging and passive infrared sensors to detect human and animal motion.

Besides, the literature commonly employs image histogram analysis for human detection [6], [7]. Filonenko et al. [8] proposed a method for smoke detection, anticipating potential explosions, using background subtraction and morphological analysis. Similarly, vehicle identification through histogram analysis is suggested in [9], while Rathor et al. [10] utilized histogram and local binary pattern analysis for animal detection.

The CV-based techniques also include human fall detection through fall motion vector analysis [11] and human gait classification via histogram analysis [12]. Video analytics engines predominantly rely on optical sensors for surveillance, although recent developments integrate infrared sensors [13], geo-location, and smoke sensors [14].

Furthermore, histogram analysis proves useful in diverse application scenarios, such as fire detection [15] and gun detection [16], with acceptable levels of accuracy.

1.2. Focus on Alert Module & Security Features

While there has been considerable attention given to the video analytics engine of a SWS, the alert module associated with it has been overlooked. Traditionally, SWSs have been equipped with non-customizable alert modules, thereby limiting their synergy with the analytics engine [1], [25]. The aforementioned SWSs [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16] are typically designed to detect objects and generate push notifications on computer screens, but they suffer from a limitation in multi-tier routing of alert signals required in a multi-level threat scenario.

There is a growing trend towards Internet-of-Things (IoT)-based alert routing. However, existing SWSs face security concerns, particularly with IoT-based alert modules being vulnerable to eavesdropping and hacking [26]. Additionally, the use of Frequency-Hopping Spread Spectrum (FHSS) communication technology in radio communication exposes SWSs to cyber threats like jamming [27].

To address these security challenges, a combination of FHSS with Advanced Encryption Standard-128 (AES-128) and Global System for Mobile Communications (GSM)-based alert provisions can enhance the security of the alert module [28]. Therefore, for a robust SWS, it is essential to incorporate multi-sensor input and implement multi-tiered routing of encrypted alert signals to ensure secure and reliable operation.

1.3. Research Questions

In the aforementioned context, our research aims to address the critical issue of unauthorized access in contemporary surveillance systems, posing a significant threat to the integrity and reliability of monitoring infrastructure. In light of this, we pose the following research question: 

How can unauthorized access to the surveillance system be prevented to ensure system integrity and reliability?

Additionally, we explore the optimal strategy for alert generation, prioritization, and secure transmission in response to potential threats, leading to our second research question: 

When a potential threat is detected, what is the optimal strategy for alert generation, prioritization, and secure transmission?

Furthermore, we investigate how the surveillance system can ensure robustness against alert signal interference, forming our third research question: 

How can the surveillance system ensure robustness against alert signal interference?

1.4. Contribution to Knowledge

To contribute to the existing body of knowledge, this paper introduces an intelligent alert module integrated into an enhanced video analytics engine for advanced SWSs. We propose a general framework (Section 2) capturing access control, multilevel threat detection, and multi-tiered alert routing, forming the basis for the proposed SWS (Section 3) using the Video Surveillance and Motion Detection (VSMD) application. VSMD, tailored for restricted areas, utilizes optical imaging, Passive Infrared (PIR) sensors for human motion tracking, and fingerprint sensors for secure access (Section 3.1). A new alert module is developed to generate multilevel alerts intelligently, routing them securely via mobile, internet-based alerts (GSM, GPRS), and encrypted Radio Frequency (RF) signals with FHSS (Section 3.2). The system seamlessly integrates this alert module with the multi-sensory analytics engine, enhancing versatility, efficacy, and security for rapid response capabilities (Section 4). The original contributions and distinctive technical advancements reported in this paper over existing solutions are as follows:

• •

  A comprehensive framework is introduced, unifying an SWS with access control, multilevel threat detection, and multi-tiered alert routing.

• •

  The innovative utilization of the VSMD application, equipped with optical imaging, PIR sensors, and fingerprint sensors, surpasses conventional binary threat detection by enabling multilevel threat assessment, access control, and enhanced security.

• •

  A novel alert module to intelligently generate threat-level-specific alerts, securely transmitted through mobile, Internet, and encrypted RF (FHSS) signals, thereby enhancing real-time threat awareness with swift response.

3.1. VSMD with an Improved Analytics Engine

As illustrated in Fig. 4, the object detection process involves gray-scale conversion, edge detection, and HOG feature extraction. To extract HOG features, pixel gradients are calculated for each 

$8\times 8$

non-overlapping pixel cell, producing a 9-bin gradient orientation histogram (each bin separated by 20 degrees). The resulting block-wise normalized gradient feature vectors, obtained from respective histograms, are then fed into an Error Correcting Output Code (ECOC) ensemble classifier for object classification. The ECOC employs multiple one-vs-one binary SVM classifiers, which are trained using a custom dataset containing images of firearms, explosions, animals, humans, and vehicles (sampled from [29], [30], [31], [32]).