Deep Learning approaches for detection and tracking of multiple moving objects in video surveillance

Abstract

In numerous real-time applications like autonomous driving, visual surveillance, sports analysis, etc, Multiple-Object Detection and Tracking application is immensely used. With the rapid development of deep learning networks, the performance of object detectors and trackers have been greatly improved. However, the existing methods lack detection and tracking due to various factors like change of appearance, complex dynamic background and occlusion. The work presented here has three major contributions to create an effective and a robust system to detect and track multiple objects in crowded scenes. The first research contribution proposes a moving multi-object detection and tracking model by exploiting a Deep Learning model along with Pearson Similarity-centered Kuhn-Munkres algorithm. The second work contributes a Leaky Multiout-Residual Neural Network classifier which executes the task of moving multi object detection and tracking along with shadow removal techniques, where shadows are often misclassified as objects and it tends to affect the classification performance. The third contribution proposes a system using Scaled Non-Monotonic Cauchy Dense Convolutional Neural Network which is an advanced technique employed in real time performance of object detection and tracking. newline newlineFor efficiently detecting and tracking the numerous objects in the complex environment, a Pearson Similarity-centered Kuhn-Munkres (PS- KM) algorithm is proposed. The input videos are initially gathered and converted into frames. The background subtraction filters the inappropriate data as of the frames after frame conversion. Then, the extraction of features is done from the frames. newline