Object Detection Using Stacked YOLOv3

Abstract

Object detection is a stimulating task in the applications of computer vision. It is gaining a lot of attention in many real-time applications such as detection of number plates of suspect cars, identifying trespassers under surveillance areas, detecting unmasked faces in security gates during the COVID-19 period, etc. Region-based Convolution Neural Networks(R-CNN), You only Look once (YOLO) based CNNs, etc., comes under Deep Learning approaches. In this proposed work, an improved stacked Yolov3 model is designed for the detection of objects by bounding boxes. Hyperparameters are tuned to get optimum performance. The proposed model evaluated using the COCO dataset, and the performance is better than other existing object detection models. Anchor boxes are used for overlapping objects. After removing all the predicted bounding boxes that have a low detection probability, bounding boxes with the highest detection probability are selected and eliminated all the bounding boxes whose Intersection Over Union value is higher than 0.4. Non-Maximal Suppression (NMS) is used to only keep the best bounding box. In this experimentation, we have tried with various range of values, but finally got better result at threshold 0.5.

Self-Correction Ship Tracking and Counting with Variable Time Window Based on YOLOv3

“…e network consists of 53 convolutional layers of 3 × 3 and 1 × 1 [21]. In response to the disappearance of gradients that may be caused by too deep layers of feature extraction, residuals are used to greatly reduce the channels of each convolution, and 3 feature images of the input with different scales are multiscale predicted to output [22]. YOLOv3 target detection algorithm is shown in Figure 1.…”

“…eoretically, when tracking a target, the direction of the movement of the target can be calculated according to the area coordinates of the same target in two adjacent frames. For example, if the coordinates of the target in the first frame are (12,22) and the coordinates in the second frame are (14,22), the target moves to the right side of the screen. However, in real situation, the coordinates of the target object may be jittered as the screen and the target move synchronously.…”

“…However, in real situation, the coordinates of the target object may be jittered as the screen and the target move synchronously. If the above method is used, the object moving in the same direction is very likely to appear at (12,22), (11,22), and (13,22), and a moving order of first to the left and then to the right may be concluded. e multiframe regression algorithm can handle this concern: regression analysis of the coordinates of the target object in several frames before and after the current frame should be conducted, the movement tendency of the object is obtained, and then a comprehensive judgment is made by combining the coordinate comparison of two adjacent frames.…”

Dual Architecture Deep Learning Based Object Detection System for Autonomous Driving

“…Using fast R-CNN the transfer learning techniques were used to generate the multi-channel images. Multi-channel networks are also used in other applications, [15] proposed a stack of YOLOv3 for mask detection in security checkpoints during Coronavirus disease (COVID-19). [16] designed a multi-class CNN to classify input images of liver lesions into sub-groupings of marginal and internal patches.…”

Differentiating Stagnant Water from Wet Surface for Detecting Potential Mosquito Breeding Sites in Real Time

“…In addition to that, we are detecting not only stagnant water but wet surface around it which makes this work novel. Therefore, a real-time object detection algorithm YOLO V3 was chosen due to its good accuracy and speed (9) . In this research, we had success in detecting various forms of water like muddy, transparent, blue, green, and black water.…”