Real-Time Object Detection With Densely Connected Network

With the development of image processing technologies,video surveillance is widely used in various scenarios of daily life,such as pedestrian detection,abnormal behavior detection,and traffic light detection.At present,the accuracy and the speed of an object detection algorithm are not balanced.As the detection accuracy increases,the detection speed decreases significantly,and it is difficult to achieve real-time detection.Therefore,high-accuracy real-time object detection has always been a hot topic in the field.Aiming at the above problems,this paper studies the existing object detection algorithms,and propose an improved algorithm which has high accuracy and high detection speed.The main work of this paper is as follows:(1)Deeply analyze the YOLOv3-TINY algorithm,and the densely connected network is intoduced to overcome its defects.A new feature extraction network is designed.The network keeps the low computation load while the number of network layers reaches 189.Due to the channel merging operation in densely connected networks,shallow features can always be transmitted in the network,the network can fully extract object features.(2)Based on the advantages of densely connected networks,multi-scale prediction is introduced,and a prediction layer of is added to improved network.Detection is performed on a feature layer with larger scale.These help to improve the results of small object detection.(3)Anchor box is obtained by clustering the training dataset and used for position prediction,which speeds up the network training and makes more accurate prediction for the object position..In this paper,PASCAL VOC and COCO datasets are used to verify the improved algorithms experimentally.Due to the densely connected network,Dense-TINY achieved a detection accuracy of 65.94% on the PASCAL VOC test set,the accuracy is11.87% higher than that of YOLOv3-TINY,and the detection speed decreased by only2 FPS.Owing to multi-scale prediction,Dense-TINY-3 has greatly improved the resultsof small object detection.The detection accuracy has been raised to 67.98%,with the detection speed 63 FPS.Using the anchor box obtained from the dataset clustering,the detection accuracy of Dense-TINY-3 reached 68.93%,0.95% higher than that without a specific anchor box.The accuracy of the object position increased by 2%.Dense-TINY achieved detection accuracy of 19.3% on the COCO dataset,which is 11.4% higher than YOLOv3-TINY.The algorithm proposed in this paper maintains a high detection accuracy,and the detection speed can still reach 63 FPS.These have shown that this proposed method has great application prospects in the field of video object detection.