Real-Time Detection of Small Targets for Video Surveillance Based on MS-YOLOv5
Pengyong Zhang, Wei Hou, Di Wu, Baoyu Ge, Lili Zhang, Huizi Li · 2023
In our work, an improved YOLOv5 model is proposed to detect small targets such as pedestrians in video surveillance in real time. We propose an improved model: MS-YOLOv5, by introducing the Transformer multi-head self-attention mechanism on the basis of YOLOv5, the deep convolutional neural network is combined with the Transformer to explore the potential of feature representation. At the same time, the Neck is added to feature up-sampling to enlarge the features of small targets, so as to improve the feature fusion structure of FPN(Feature Pyramid Networks)+PAN(Path-Aggregation Network) and further improve the detection precision of small targets. After the improvement of the above method, YOLOv5 was finally named MS-YOLOv5. By training the MS-YOLOv5 model using small targets in video surveillance, detection experiments are carried out for the main types of targets appearing in video surveillance. The MS-YOLOv5 detection results were compared with the original YOLOv5 detection results, and the precision was increased by 13.7% and the recall was increased by 11.3%, and the MS-YOLOv5 model significantly improved the detection precision of small targets and reduced the missed detection rate of small targets.