Efficient Change Detectors for Intelligent Video Analytics

Abstract

Detecting an intruder that is trespassing a prohibited area is a critical task of intelligent video analytics. This task requires a change detector to segment an intruder (foreground object) from the background. The task suffers the inherent drawbacks of change detectors due to the dual camera sensors (color/IR), illumination changes, night time, static, and camouflaged foreground objects. This work proposes efficient unsupervised and supervised change detectors to compensate for the aforementioned challenges for intelligent video analytics particularly industrial sterile zone monitoring.

The camera switch detection based on skewness patterns detects a switch between the dual camera sensors (color/IR). The optimal color space selection based on the mean squared error will select tolerant color space (RGB/YCbCr) to illumination changes for modeling the background. Also, the IR camera frames are contrast-enhanced to tackle the camouflaged intruders during the night. The incoming frames are split into respective channels before modeling the background. The background is modeled by Gaussian Mixture Models (GMM). The adaptive background model update scheme is proposed to tackle the various challenges posed by environment and object such as a static foreground object. Convolutional Neural Network (CNN) based algorithms have shown promise in dealing with the aforementioned challenges. However, they are exclusively focused on accuracy. This work goes on proposing an efficient supervised change detection algorithm based on atrous deep spatial features. The features are extracted using atrous convolution kernels to enlarge the field-of-view (FOV) of a kernel mask, thereby encoding rich context features without increasing the number of parameters. The network further benefits from a residual dense block strategy that mixes the mid- and high-level features to retain the foreground information lost in low-resolution high-level features. The extracted features are expanded using a novel pyramid upsampling network. The feature maps are upsampled using bilinear interpolation and pass through a 3x3 convolutional kernel. The expanded feature maps are concatenated with the corresponding mid and low-level feature maps from an atrous feature extractor to further refine the expanded feature maps. The experiments were performed on three standard change detection and video analytics databases. The proposed algorithms showed better performance than high-ranked unsupervised and supervised change detection counterparts on the three standard databases.