Research And Implementation Of Pedestrian Detection System Based On Deep Learning

In recent years,with the rapid development of computer vision technology,pedestrian detection,as an important research field,has also made great progress and is gradually applied to practical applications.However,the existing traditional detection algorithms still have problems such as partial feature extraction,low detection accuracy,and high time complexity.With the research and application of deep learning algorithm,a series of deep learning detection algorithms are derived based on convolutional neural network.Compared with traditional detection algorithms,deep learning algorithm has stronger robustness and generalization ability,and can detect pedestrian targets faster and more accurately.Thanks to the continuous innovation and optimization of pedestrian detection theory,pedestrian detection provides technical support for intelligent monitoring,unmanned driving and other aspects,and has great application value.However,in the actual monitoring scene,the current pedestrian detection algorithm still has the problem of pedestrian misdetection,and is easily affected by the factors such as occlusion,pedestrian posture and scale change.The detection performance needs to be further strengthened.In order to improve the pedestrian detection effect in actual monitoring scene,this paper comprehensively analyzes the classic network detection model,and selects the SSD network model with good detection accuracy and real-time performance as the foundation,optimizes and improves the problems of missed detection,false detection,and time-consuming,etc..An improved SSD network that makes full use of shallow feature information is proposed,and a pedestrian detection system is designed on the Jetson Nano Developer Kit.The main work is as follows:Firstly,the basic SSD model network model and training process are studied,and the Py Torch framework is used to build a pedestrian detection network for pedestrian detection experiments in INRIA and Caltech data sets.According to the experimental results,the shortcomings of the SSD algorithm are analyzed and provide comparison benchmarks for improved algorithms.Then,the existing problems in SSD are optimized,and the improved algorithm uses Res Net residual network integrating shallow feature information to replace the traditional VGG network,and adds Vo VNet on this basis to reduce network parameters and calculation amount through the OSA module.At the same time,the K-means clustering algorithm is used to optimize the anchor setting.Use the improved SSD and SSD algorithm to conduct a comparative experiment,the experimental results show that the improved SSD algorithm can solve the problems of missed and false detection and long time consumption caused by dense or occluded pedestrians and small pedestrian posture.Finally,on the basis of the above research and improved algorithm,a pedestrian detection system was built based on Jetson Nano Developer Kit,and the improved SSD and SSD algorithm were compared for comparison experiments.In addition,a separate label detection experiment was performed on the collected data set;the client of the pedestrian detection system is built,and the improved algorithm is used to detect pedestrians,and the pedestrian images are saved for later retrieval,which realized the mobile terminal pedestrian detection.