Research On Precise Pedestrian Detection Based On Deep Learning Autonomous Driving

In recent years,with the rapid development of deep learning technology and high-performance computers,the target detection method based on machine vision has penetrated into various fields.Pedestrian detection technology is one of the popular research directions in the field of computer vision and plays an important role in monitoring security and automatic driving of cars.Especially in the aspect of automatic driving,road pedestrian detection is one of the most important detection and analysis targets in automatic driving application scenarios,and accurate detection of pedestrians in the environment is an important prerequisite for completing subsequent tasks or human-computer interaction.Compared with traditional pedestrian detection technology,GPU-based deep learning pedestrian detection technology has great advantages in detection accuracy and detection speed.At the same time,with the rapid development of edge computing technology based on embedded platform in the field of autonomous driving,the deployment of network model on embedded terminal for real-time pedestrian detection has also become a research hotspot and difficulty.In this paper,the RetinaNet network based on the one-stage method is studied and improved.Due to the limited computing resources of the embedded platform,the one-stage method is more suitable for the target detection network with lower requirements on computing resources.This paper will improve the embedded platform of autonomous driving.There are two aspects to the improvement of RetinaNet network in this paper.In this part,this paper compares different lightweight methods,and introduces the MobileNet network to lightweight the feature extraction network,so that the network model has more pertinence and adaptability to the embedded platform.Second,the loss function of the classification subnetwork is optimized,and the L2 softmax loss function is introduced.This loss function has certain advantages in the processing method of pedestrian characteristics,which can better improve the robustness of the network,and to some extent makes up for the decrease of accuracy caused by the lightweight treatment of the network.In order to verify the effectiveness of the modified network,the network is trained in the experimental part and compared with other networks.Before the experiment,the dataset was firstly processed.In this paper,the COCO dataset was selected,screened and processed with data enhancement to improve the robustness and generalization ability of the network.The pre-training method is used to accelerate the convergence of the network.After training,the results show that the network reduces the number of parameters to about 30%of that of the original network under the condition that the accuracy is basically unchanged,and takes into account the requirements of low number of parameters and high recognition accuracy.At the same time,the improved network was compared with the common target detection network,and the training was carried out on the dataset of this paper and the VOC 2012 dataset respectively.The results show that the improved model has better comprehensive performance and is more suitable for application in embedded platform.It can better enable the automatic driving system to make timely and accurate judgment,and provide prerequisite guarantee for subsequent corresponding driving strategies and protection measures.