Deep Learning-based Violation Detection Of Persons In Oil Deport

In recent years,the security problems of oil depot occasionally occured,and it is increasingly important to pay attention to the violations of oil depot staff.However,it is obviously inadvisable to monitor the cameras manually.Therefore,object detection based on deep learning is of great significance.A Faster RCNN model was constructed in this paper to detect whether persons in oil deport wear helmets or tooling suits.The construction is divided into three steps:(1)The videos of seven oil depots were collected from the internship company.And the key frames containing targets were extracted by combining random interception and frame difference method.The video frames were flipped 180°horizontally to increase data richness;(2)The extracted video data were marked into four categories: {wearing tooling and helmet,wearing tooling without helmet,wearing helmet without tooling,and wearing neither tooling nor helmet};(3)The improved Faster RCNN was used for training and testing.The main improvements of Faster RCNN were as follows: the training images were trained by multiscale random selection method.The short edge of images is randomly selected from 480,600 and 750 while the original proportion of images remains unchanged to avoid low accuracy.Then the resized images are sent to the network for training to avoid low accuracy or missed detection due to different target sizes;ResNet-50 is used to replace the traditional VGG16 for feature extraction,which greatly improves the feature extraction ability of the model and avoids the problems caused by the deepening of network depth;In the RPN phase,in addition to generating the nine default Anchors,another 3 Anchors(24×24,64×64,128×128)are generated as a fixed anchor scale so that the improved RPN can almost cover all objectives.In order to verify the effectiveness and detection accuracy of the improved Faster RCNN method,evaluation indexes mAP and MR-2 were constructed.Experiments are carried out on the collected data set in this paper.When the coverage threshold is 0.5,the mAP value raised from 84.13%(traditional method)to 96.63%(improved method).The detection accuracy was significantly improved 12.5%.The MR-2 decreased from17%,23%,33% and 25% to 6%,12%,11% and 13% respectively,indicating that the missed rate has been significantly improved.Since there is no obvious correlation between wearing helmet and wearing tooling,data sets of helmet and tooling were constructed respectively.The mAP value of helmet data set increases from 85.06% to98.51%,and that of tooling data set increases from 87.01% to 97.87%.The results show that the improved Faster RCNN is effective for helmet and tooling detection,and the target detection method proposed in this paper has certain research and application value.