Research On Pedestrian Detection Algorithm Based On Deep Learning And Its Deployment And Application In Embedded System

Pedestrian detection refers to the use of rectangular boxes to accurately mark the pedestrians in pictures or video sequences.It plays an important role in many fields,such as robots,intelligent driving and safety monitoring.With the rapid development of convolutional neural network and computer technology,pedestrian detection method based on deep learning has occupied a dominant position and has been widely used.This paper mainly studies with the help of SSD and Yolo V4 algorithm to improve the detection accuracy and speed of the algorithm and complete the deployment on the embedded platform.SSD(Single Shot Multi-Box Detector)is a neural network algorithm widely used in pedestrian detection.In order to improve its detection accuracy and detection speed,the SSD has been effectively improved(the improved algorithm is called XSSD-P).First,by comparing the feature extraction ability and speed of different backbone networks,the Xception network is selected as the backbone network of XSSD-P,and the feature layer for prediction is re-selected through experiments.According to the appearance characteristics of pedestrian,multi-scale convolution kernels and basic anchors are designed,and the two are coupled.The basic anchors slides on the input image following the multi-scale convolution kernels,and anchors are obtained by adjusting its size for position regression.Finally,the depth separable convolution with less parameters and calculation is used to predict on the feature map instead of the conventional convolution.The above improvement realizes the effective detection of pedestrians.Compared with SSD and other mainstream algorithms,XSSD-P has higher detection accuracy in pedestrian detection.And in terms of detection speed,compared with the SSD,the detection speed of the XSSD-P is 30 FPS higher,an increase of42.86%.Experimental results show that the detection accuracy and detection speed of XSSD-P are better than those of SSD.Although the detection accuracy of the current pedestrian detection algorithm has been greatly improved,there is still a serious problem of pedestrian missing detection in the case of pedestrian occlusion or small pedestrian size.Aiming at this problem,this paper first analyzes the impact of feature loss and NMS on the missing detection of YOLO V4,and comes to the conclusion that feature loss is the key cause of pedestrian missing detection.In order to alleviate the influence caused by the lack of features,this paper establishes a pedestrian local detector to detect the body parts of pedestrians,and then establishes the size and position relationship between the pedestrian part and the whole pedestrian through regression analysis and machine learning,so as to obtain the position information of the whole pedestrian according to the position information of the pedestrian part.Finally,the comparative test in the test data set shows that the recall rate of the improved algorithm is greatly improved.The test results in the data set show that the maximum improvement is 8.62%.In order to meet the needs of practical application,NVIDIA Jeston Xavier NX core modules and related software are used to build an embedded platform and deploy the model.Firstly,the training model obtained by the training framework is transformed into ONNX intermediate format with the help of Python library,and then the model deployment is completed by using tensorrt reasoning framework and its library function.Tensorrt reasoning framework effectively reduces the reasoning speed of the model by reducing the calculation accuracy of the model and reconstructing the network structure of the model.The effectiveness of the method is verified by the test of data set and actual scene.Finally,the practical application of the improved model is realized.