Deep learning based methods for improving object detection classification and tracking in video surveillance

Abstract

In computer vision, object detection, classification, and tracking newlinewere considered as the challenges which are applied in a variety of fields that newlineinclude video surveillance, autonomous vehicles, and human-machine newlineinterfaces. The one-stage and two-stage methods are the core types of newlineadvanced object detection techniques. However, the performance in terms of newlineaccuracy and effectiveness is not highly satisfactory. The present work dealt newlinewith detection, classification, and tracking objects in surveillance video. The newlinefirst research contribution dealt with detection and classification of objects newlineusing the Modified Manta-Ray Foraging Optimization-based Convolution newlineNeural Network (M2RFO-CNN) in a complex dynamic environment. To newlineachieve higher accuracy, the second work employed Batch normalization and newlineSoftswish activation adapted ResNet (BS2ResNet) and Logistic Tanh newlineKaiming Bi-directional Long Short Term Memory (LTK-Bi-LSTM) for newlinedetection and classification, respectively. Efficient object detection and newlinetracking in adverse weather conditions using Tanh Softmax adapted newlineEfficientDet (TSM-EfficientDet) and Jaccard Similarity-centric Kuhn- newlineMunkres (JS-KM) integrated with Pearson-Retinex was the third major newlinecontribution of the present study. The existing models suffer from being unable to provide high accuracy and precision in detecting dense and smaller objects in the complex newlinedynamic environment. In the first proposed work, in the resized noiseremoved newlineframes, Contrast Limited Adaptive Edge preserving Algorithm newline(CLAHE) was applied to boost the contrast. The contrast-enhanced frames newlinewere trained using HTYOLOV4. Initially, the frames were split into grids. newlineHowever, the inappropriate size of grids can result in inaccurate detection. To newlineIv resolve this issue, the Hyperbolic Tangent Kernel function was utilized to newlineoptimize the grid levels by minimizing the loss function in YOLOV4. The newlineGrasp configuration was extracted for every bounding box in improving the newlinedetection accuracy of smaller objects. Values below the threshold were newlinedeemed