An Effective Object Tracking Using Yolov3 with Bidirectional Feature Pyramid Network on Video Surveillance

Deep learning enhances precision and efficiency in object tracking, enabling applications in autonomous vehicles and surveillance while adapting to dynamic environments. To address issues such as overlapping and anchor box regions of interest for tiny objects, this research introduces a YOLOv3 model with a Bidirectional Feature Pyramid Network (BiFPN) for overcome the start-of-art methods problems. The FLIR dataset is employed for effective object tracking and classification, with preprocessing techniques like Histogram Equalization and Laplacian sharpening to reduce Gaussian noise and enhance object edges in the benchmark dataset. The pre-processed images and video clips are input into the YOLOv3 and BiFPN for detecting both tiny and major objects and constructing anchor boxes. The results demonstrate the outstanding performance of the YOLOv3 with BiFPN model, achieving an impressive mean average precision (mAP) of 90.50%. It also offers rapid detection, with a time of only 6 milliseconds, and maintains a remarkable frames per second (FPS) value of 5, ensuring superior classification compared to other existing methods such as NS EfficientDet and Tiny YOLOv3 network.