Research On Small Target Detection Algorithm Based On Multi-scale Feature Fusion

As more and more people choose to travel by subway,in order to ensure the safety of passengers,real-time and accurate monitoring of passenger flow in subway stations has become particularly important.Different from other open-air scenes,there is a large passenger flow in subway stations during peak hours.At the same time,the high installation location of video capture equipment in the station and the problem of video capture angles cause the pedestrian targets in video surveillance to have small pixels and difficult detection.This article proposes Multi-scale weighted feature fusion network realizes accurate real-time monitoring of subway passenger flow.The main contributions include the following:1.Data preprocessing: This paper proposes a small target enhancement algorithm based on oversampling.In the entire data set,there are very few pictures containing small targets,but almost all targets in individual images are small targets.This unbalanced data distribution seriously hinders the training process.The small target enhancement algorithm based on oversampling effectively adds small target instances in the data set,and increases the iterative frequency of small target instances in the training process during the training process,and increases the weight contribution of small targets.2.In crowded scenes such as subway stations,due to problems such as high pedestrian density and severe occlusion,most existing pedestrian detection methods are difficult to obtain good results and are less robust to small targets.In response to this problem,this paper proposes an improved human body detection network model based on SSD(Single Shot Multi Box Detector).First,analyze the characteristics of subway pedestrian data and briefly describe the characteristics of the head-and-shoulders detection model.Then,the SF-SSD(Supplementary feature single short multibox detector)network was proposed.Based on the SSD network,the low-level feature maps are merged,and their location information is used to accurately detect small targets.Finally,experiments were conducted on the subway pedestrian dataset.The experimental results show that the proposed SF-SSD network method effectively reduces the false alarm rate and missed detection rate of the SSD network on the subway pedestrian data set,meets the real-time requirements of pedestrian detection,and achieves accurate detection in dense scenarios.3.Taking the SSD network as the basic model,it has been further improved,and a feature extraction layer based on the VGG16 network is added,and a large number of experiments are performed to determine the optimal parameters to perform multi-layer and multi-scale feature fusion.Finally,experiments are performed on the data processed by the small target oversampling enhancement algorithm,and a multi-scale weighted feature fusion model is obtained.Secondly,for the problem of feature channel weight distribution,a multi-scale adaptive channel attention mechanism model MWFSSD is proposed to realize automatic learning of weight parameters and improve the detection effect without affecting the real-time performance.Experiments have proved that the improved method proposed in this paper has a 6.82% increase in detection accuracy compared with the detection effect of the original SSD,and can be effectively applied to passenger flow detection in subway stations.