Published Paper Details:

In the context of computer vision, object tracking is a crucial challenge. Image noise, variations in backdrop illumination, dynamic object motion, and partial and complete occlusion of objects all add to the complexity of object tracking. i.e. video of poor quality. Despite the impressive outcomes of deep learning, there has been less emphasis on low-resolution object recognition as compared to high-resolution images. We proposed a framework for reducing model size state-of-the-art single stage detector i.e. YOLOv4 with pruning and Kalman filtering for tracking the low quality video. The framework addresses compression in the following stages: Sparsity Induction, Filter Choice, and Filter Pruning. The detector model for detecting object is sparsified in the Sparsity by using a better global threshold. In the Filter Selection & Pruning stage, we choose and remove filters based on sparsity metrics of filter weights in two subsequent convolutional layers. As a result, the model is smaller than the majority of compact designs available now on the market. We test our framework's performance with a MSCOCO datasets. The model with compressing algorithm with having Kalman filter to produce good quality image representation on the low-quality input image or video feed analysis proposed for the condition. The models size examines for the inferences and the analytical framework for the better represent for the detection stages and evaluation results.