Research And Implementation Of Human Pose Estimation Algorithm In Autonomous Driving Scenario

Pedestrian safety has always been a key issue in the driving field,and it has also become one of the key factors restricting the development of autonomous driving.Early methods dealt with pedestrians as special obstacles,ignoring the movement characteristics of the human body.With the rapid development of deep learning in the field of human pose estimation,more and more models with high accuracy and high detection speed have been proposed one after another,and its application to autonomous driving decision systems has become a hot research direction.This article first reviews the current mainstream 2D human pose estimation methods,and then uses deep learning methods based on street view images taken by monocular cameras to explore the feasibility of human pose estimation in an autonomous driving environment for more demanding pedestrians safety provides theoretical help.Aiming at the long-distance low-resolution street scene images obtained by the vehicle camera,this paper uses a bottom-up design idea to construct a multi-person human pose estimation network model.Compared with traditional feature-based feature extractors,deep convolutional neural networks can automatically learn the optimal image features.In this paper,the receptive field is increased by accumulating convolutional layers,so that the network model can learn more global and higher semantic features,achieve the purpose of implicitly learning human body structure information,and enhance the generalization and robustness of network models.At the same time,the relay function is used to construct the loss function to prevent the disappearance of the gradient caused by the network being too deep.The method achieved 72%and 68%accuracy on the benchmark data sets MP? and MS COCO respectively,far exceeding the detection accuracy of traditional algorithms.Aiming at the problem that the network model has a weak detection ability in outdoor scenes,this paper introduces a dynamic residual structure based on the original network structure to enhance the network's ability to identify features so that the model can adapt to different scenarios.At the same time,a special out-of-vehicle scene dataset was added on the basis of the benchmark dataset,and the key points of the human body in the scene were labeled in two dimensions,and the training set was enhanced by digital image processing technology.The enhanced network model's detection capability in outdoor scenes is 2 percentage points higher than the original model.Aiming at the difficulty of directly deploying the network model to a vehicle-mounted terminal with limited hardware resources,the network model is deeply optimized.In this paper,a deep separable convolution is used instead of the standard convolution operation,which reduces the model parameters while ensuring a large receptive field and greatly reduces the amount of convolution calculations.In addition to improving the Backbone network,this paper also streamlines the RefineNet structure by merging redundant parallel branches and using residual blocks instead of large convolution kernels.Experiments show that the final optimized network model reduces the volume of the entlire model by 83%while ensuring approximate accuracy,and the detection effciency is increased by ten times.The human pose estimation network model constructed in this paper is expected to be applied to the field of autonomous driving.By detecting human skeleton lines to analyze pedestrian behavior and predict pedestrian walking trajectories,safety warnings can be made in advance,which can greatly improve driving safety.