Adaptive Pixel Deviation Absorption Technique for Efficient Video Surveillance using Deep Convolutional Neural Networks

Monitoring human activity in industries is a great challenge and numerous methods use various features, such as sketch, position, color, and shape features. However, these methods do not achieve the expected accuracy in classifying the activity of people in the environment. This study presents an Adaptive Pixel Deviation Approximation with Deep Convolutional Neural Networks (APDA-DCNN) model to increase classification accuracy. The method starts with local feature-approximation-based normalization of video frames. Then, global value segmentation is used to group the features of the frame. From the image segmented, the human features along with texture and region pixel deviation features are extracted. The APDA-DCNN model trains the CNN model to convolve the texture features into one-dimensional features by convolving in two layers. The output layer neurons estimate Texture Similarity (TS), Sketch Level Similarity (SLS,) and Pixel Deviation Similarity (PDS) against various classes. Using the values of TS and PDS, the model estimates the Activity Weight (AW) against various classes to select the most dominant. The APDA-DCNN model increases the accuracy of activity classification to achieve higher video surveillance performance.