Change Detection Algorithms for Surveillance in Visual IoT: A Comparative Study Visual Internet of Things

The VIoT (Visual Internet of Things) connects virtual information world with real world objects using sensors and pervasive computing. For video surveillance in VIoT, ChD (Change Detection) is a critical component. ChD algorithms identify regions of change in multiple images of the same scene recorded at different time intervals for video surveillance. This paper presents performance comparison of histogram thresholding and classification ChD algorithms using quantitative measures for video surveillance in VIoT based on salient features of datasets. The thresholding algorithms Otsu, Kapur, Rosin and classification methods k-means, EM (Expectation Maximization) were simulated in MATLAB using diverse datasets. For performance evaluation, the quantitative measures used include OSR (Overall Success Rate), YC (Yule’s Coefficient) and JC (Jaccard’s Coefficient), execution time and memory consumption. Experimental results showed that Kapur’s algorithm performed better for both indoor and outdoor environments with illumination changes, shadowing and medium to fast moving objects. However, it reflected degraded performance for small object size with minor changes. Otsu algorithm showed better results for indoor environments with slow to medium changes and nomadic object mobility. k-means showed good results in indoor environment with small object size producing slow change, no shadowing and scarce illumination changes.


INTRODUCTION
T he term IoT (Internet of Things) was first used by Lee et. al. [1] defined IoT, also called the Internet of Everything or the Industrial Internet, "is a new technology paradigm envisioned as a global network of machines and devices capable of interacting shopping and surveillance [2]. VIoT deals with image and video processing in particular [3]. In recent years, IoT and VIoT have garnered huge research interest due to evolving and merging heterogeneous technologies.
Surveillance refers to monitoring of behavior, objects and events whereas DVS (Digital Video Surveillance) is simply surveillance conducted using cameras. DVS has found widespread role in security [4] [5] or VIoT containing heterogeneous devices with varying computational and storage capabilities. These devices may include resource constrained visual sensor nodes, with lower cost, smaller size, limited processing power, average quality cameras and smart phones cameras [6]. Computationally powerful device such as gateway is used to relay the information to Back Office. For lowform factor VIoT devices, it is required to employ techniques which pose less computational and communication overhead [7].
A common DVS for VIoT comprises several phases as shown in Fig. 1. ChD is performed as in-network processing on low-form factor VIoT devices. For video surveillance, object classification and tracking is done as Back-office processing on servers.
The efficiency and effectiveness of VIoT based DVS is dependent upon algorithms used at every step. This paper focuses on comparison of various ChD algorithms suitable for resource constrained devices in VIoT.
ChD refers to monitoring temporal changes in images.
ChD algorithms can be categorized in several ways. One way to categorize them is based on image representation [8] namely pixel based, region based and edge based.
Another categorization is based on algorithm being supervised or unsupervised. ChD algorithms can employ different techniques. Examples include image differencing, image ratioing, image regression, principle component analysis and component vector analysis [9]. Image differencing and rationing are computationally inexpensive and can easily be implemented on low-form factor devices with minimal resources which makes these techniques well suited for VIoT and VWSN.
For VIoT devices, ChD must be implemented using computationally inexpensive techniques to conserve memory, CPU (Central Processing Unit) cycles and battery power. Consider a typical VIoT based DVS system, where multiple VIoT devices such as visual sensor nodes, smart cameras acquire images and execute ChD algorithm. If change is detected, data is transmitted to gateway. The gateway forwards data to Back office for further processing. This process is represented in Fig. 2.

FIG. 1. STEPS EMPLOYED IN VISUAL SURVEILLANCE FOR VIOT
ChD algorithms have been compared previously as well [10][11]. However, as far as the authors know, there is no reported work that compares and suggests appropriate ChD algorithms for VIoT based specific environment settings. The scenarios where slow change and/or rapid change and illumination changes in indoor and/or outdoor are present. This work simulates multiple ChD algorithms and evaluates the performance of each in different environment settings using quantitative measures such as OSR, JC and YC. The algorithms that have been compared are pixel-based, unsupervised using image differencing. The algorithms list includes histogram thresholding algorithms Otsu, Rosin, Kapur and classification methods k-Means and EM [10,12].
The remainder of the manuscript is organized as follows; Section 2 provides a detailed review of related work; Section 3 describes the simulation framework, evaluation parameters along with results and performance analysis.  Aach et. al. [24] proposed and evaluated statistical model based video change detection targeted on noise handling.

PERFORMANCE EVALUATION
Performance of ChD algorithms can be evaluated qualitatively and quantitatively based on the application requirements. For qualitative evaluation, the user is displayed images successively while the 'change mask' produced by the ChD algorithm can be superimposed onto the image to make the change more visible.
Furthermore, color codes may be used to differentiate between original image and change image. As for quantitative evaluation, it is based on the comparison of the output produced by ChD algorithm in relation to the 'ground truth'. The ground truth is an application specific reference image which gives the exact output as required in a particular scenario [12].

Performance Evaluation Measures
Confusion matrix for analysis of English alphabets was first proposed by Townsend [29] and its usage in machine learning was popularized by Provost et. al. [30]. Confusion matrix construction has become a de facto standard to use in image processing for result categorization as well. In order to measure effectiveness and efficiency of ChD algorithms, the following measures were selected as shown in Table 3.
The suitability of these evaluation measures is described in Table 4

SIMULATION
The following histogram thresholding and classification algorithms for ChD were compared: Otsu [32], Kapur [33], Rosin [34], k-means and EM. For simulation, hardware and software specifications are summarized in Table 5.

Results Discussion
From Fig. 6 With the emergence of middleware for VIoTs, the scope of the study can be extended to solve the ChD problem.
As ANN (Artificial Neural Networks), principal component analysis, and change vector analysis can be run at the BackOffice to explore further possibilities.
Relationship among OSR, YC and JC can be further studied to gain more insight in the results obtained along with inclusion of other accuracy measures as suggested for better evaluation of ChD algorithms.