COA-SSD: A novel chaotic optimization algorithm for enhancement of real-time video surveillance images using single shot detector with MobileNetV2

Video surveillance involves the deployment of cameras to observe and record activities for security and moni
toring purposes. This is vital for crime prevention, public safety, and improving operational efficiency across
various environments. The rising demand for real-time security and threat detection has increased interest in
sophisticated video surveillance systems. However, challenges persist in achieving an accurate automated
analysis under diverse environmental and situational conditions. This paper presents an optimization strategy for
video surveillance system deployment using a deep-learning algorithm. This strategy combines the base image
with information about the behavioral events depicted in the frame sequence and then uses a Sobel filter to
detect boundaries. The present study employs a rapid feature point identification algorithm known as the local-
peak scale-invariant feature transform (LP-SIFT), which is based on multi-scale local peak features that are
invariant to scale and represent multi-scale fluid turbulence. Traditional segmentation methods in video sur
veillance often lack contextual understanding and temporal consistency, which Modulated Memory Networks
aim to address using adaptive memory-driven feature modulation. Next, we used an (SSD) with MobileNet V2 to
classify fragments. SSDs can bypass the region proposal network and reduce latency. SSDs incorporate several
optimizations such as multiple cooperative functions and base boxes to compensate for the loss of accuracy. To
optimize and enhance the accuracy and efficiency of the video surveillance classification, the model parameters
were fine-tuned. The Coati Optimization Algorithm (COA) is a metaheuristic algorithm inspired by coati be
haviors: their cooperative strategy when attacking iguanas, and their anti-predator and avoidance behavior
strategies. To evaluate the LP-SIFT algorithm, we considered datasets with varying scenes, pixel sizes, and
distortion levels, across different lighting conditions and environments