| Object detection is the use of computer technology to mark objects in pictures or videos,and it can be seen everywhere in people’s daily life today.Target detection technology has a wide range of application scenarios.In the field of public security,police officers can quickly check the face information in the surveillance video before and after the incident,and compare it with the security blacklist personnel to quickly solve the case and improve social stability;in the guard system,Automatic license plate detection can be used to realize fast vehicle entry and exit,saving labor costs in parking lots.In recent years,the accident rate of housing and municipal engineering is on the rise,and ensuring the personal safety of construction workers is the top priority.At present,many construction sites are equipped with helmet detection systems.By configuring cameras to obtain on-site surveillance video,and using detection models to analyze and identify the video,it can quickly detect violations in the scene.Thereby reducing the potential safety hazards at the construction site and ensuring the safety of construction operations.Many current studies can effectively identify whether workers wear helmets in simple detection scenarios,but there are deficiencies in dense crowds or complex scenarios.In this paper,the YOLOX model is used as the benchmark model.First,a series of improvement measures are made to solve the problem that the model is not sensitive enough to dense target detection and is prone to false detection in the face of complex scenes.Secondly,the model has many parameters and low hardware performance in actual deployment.The problem that the platform cannot configure large-scale detection models has been improved with lightweight.The main contributions of this paper are as follows:1.Aiming at the problem that the model is not sensitive enough to dense target detection and prone to misdetection in the face of complex scenes,this paper proposes a custom attention module based on the channel attention mechanism and spatial attention mechanism,and embeds the feature fusion network of the original YOLOX.The improved attention module improves the model’s ability to perceive edge information of dense objects by further fusing the original feature map with the spatial attention feature map.In order to further improve the detection ability of the model,this paper draws on the idea of residual network and proposes the Res-FPN structure in the feature fusion network to improve the fusion of global feature information and local feature information.Finally,it is verified on the SHWD helmett dataset.Compared with the benchmark model YOLOX,the improved model has a stronger ability to detect dense targets and is less prone to false detection.The effect of each improved module is verified through ablation experiments,and the improvement of the detection effect of the overall model is verified through comparison experiments with other models.2.Aiming at the problem that many model parameters require a large amount of calculation and a large amount of resources,this paper first adopts the method of replacing the backbone feature extraction network,replacing the YOLOX backbone network from CSPDarknet-53 to Mobile Netv3 to reduce the amount of model parameters.Then,channel pruning is used to cut redundant channels with different thresholds by using the scaling factor of the batch normalization layer to reduce the amount of parameters.Finally,knowledge distillation is used to recover the accuracy of the pruned model.Finally,it is verified on the SHWD helmet dataset.Compared with the benchmark model YOLOX,the number of model parameters compressed using the lightweight strategy is significantly reduced,and the inference speed is greatly improved.The influence of different pruning thresholds on the detection accuracy and parameter quantity of the model is verified by ablation experiments.The experimental results show that the improved model is suitable for the helmet detection task in the construction site. |
PDF Full Text Request