Research On Pedestrian Detection Based On Super Resolution And YOLOv5

At present,the research field of computer vision is increasingly expanded.Its target detection technology has so far in the range of more than 90%of various fields,especially in the field of road traffic.In the field of road traffic,pedestrians occupy an important role and are the main participants.Therefore,the exploration of pedestrian testing is particularly important,and it plays a key role in the development of intelligent transportation in various cities.Traditional pedestrian testing methods are pedestrian testing based on handmade characteristics.When there are a lot of targets,hand-extruding timely time is longer,and the key points and extraction difficulties are not available.Therefore deadlock.Under the premise of convolutional neural network,there are problems of blurred image and low detection accuracy in pedestrian detection.Therefore,this paper proposes pedestrian detection based on SRGAN and YOLOv5 to study how to improve the detection accuracy and speed.The main tasks of this paper are:(1)Aiming at the problem that the captured images are blurred due to the performance of monitoring or shooting equipment in road traffic and the accuracy of image detection results is not high,an improved SRGAN is proposed.research on image preprocessing.Perform image super-resolution reconstruction on the image to be detected,use the Leaky Re LU module instead of the Re LU module in the generative model G and remove the BN layer to form a dense residual block,and use the global average pooling in the recognition model D instead of the fully connected layer to calculate each layer The average value of the feature image pixels,this improvement alleviates the gradient disappearance and overfitting to a certain extent.(2)Aiming at the problems of low detection accuracy and low detection efficiency,a pedestrian detection research based on improved YOLOv5 is proposed.Through the learning and analysis of the YOLO model,the Yolov5 algorithm was improved,and a new YOLO model idea was designed.First,the Transformer encoder block was used in the Back Bone part to replace the CSP bottleneck block and convolution block in the CSPDark Net53 network,and then the volume was inserted in the Neck part.Finally,the super-resolution reconstructed Pascal VOC 2007 dataset images are used for pedestrian detection.Through the experiments of this improved model on the dataset,the training efficiency and performance of pedestrian detection are improved compared with traditional methods.Figure 31 table 10 reference 71...