Innovative Object Detection with YOLOv4 and Adaptive Bounding Boxes

This research work proposes a computer vision-based object detection system. It can recognize each and every object visible in the picture. Horizontal boundary boxes were used to detect different objects available in the detectors. However, buildings and cars in the aerial imagery are occasionally very close to form continuous objects that may suggest a direction. Therefore, some approaches such as extending the horizontal bounding box to oriented box are proposed for a better object extraction; usually via directly regressing the angle or corners. Here we propose an orientation object detector which is simple but more accurate, and it is based on YOLOv4. Here yolov4 is used to detect objects yolo can able to find more than one object in a single image. The paper will contain specific implementation methodology for yolov4 with OpenCV, which includes model configuration settings, preprocessing steps and evaluation matrices such as mean average precision (map), recall and the precision. The object detection problem has transformed into a key problem in computer vision and has enabled improvements in areas including self-driving cars, medical systems, and even surveillance systems. In this paper, we present a new object detection system based on YOLOv4 which stands for ‘You Only Look Once’ version four and uses adaptive bounding boxes. The CSPDarknet-53 and spatial pyramid pooling (SPP) are the main reasons for YOLOv4's immense popularity, as is discussed in the literature. But with the constant changes of the bounding boxes, performance accuracy suffers greatly as they have a difficult time identifying the objects bounding boxes are placed within. To counter this problem, we present an explainable adaptive bounding box technique which relies heavily on the size shape and features of the context of the object being captured. The learnable bounding box adjustment layer which serves to refine the localized bounding boxes is integrated into the head of YOLOv4, where putting the primitive guesses goes through our specialized deep neural network so filters are set and only the most accurate localization guesses are retained. This method proves useful in more complicated scenarios with the presence of occlusions, multiple objects and changes in scale.