Research On Pedestrian Detection Algorithm Based On Deep Learning

Pedestrian detection technology is widely used in the fields of intelligent monitoring,intelligent robots and unmanned driving.With the development of agricultural equipment in the direction of intelligence and unmanned,pedestrian detection technology provides technical support for obstacle detection,obstacle avoidance and path planning of unmanned agricultural machinery operations to ensure personal safety.In this paper,the deep learning method is applied to the pedestrian detection algorithm,with a view to further applying it to the orchard environment with complex background and difficult detection.The main work of this article is as follows:?1?In view of the problems that the pedestrians in the orchard environment are blocked by fruit trees,often in motion and non-standard postures,which affect the detection performance of the algorithm,the following measures was taken for the YOLOv3 pedestrian detection algorithm to improve its detection performance:adopted the ASFF strategy to solve the inconsistency of the feature pyramid by learning the connections between different feature maps,so as to improve the scale invariance of features and the detection accuracy of the network model;used the CloU loss function to perform boundary box regression to locate the target,accelerate the convergence speed of the network model and improve the accuracy of target positioning;used the DloU NMS algorithm to improve the NMS algorithm.The DloU NMS algorithm suppression criterion takes into account the overlapping area of the prediction box and the distance between the center points of the two boxs,which can avoid false suppression of blocked pedestrians in the orchard environment.Using the NREC dataset for experiments,the results showed that the improved algorithm improved the accuracy of YOLOv3 by 2.85 percentage points to 97.63%,the recall rate increased by 3.02 percentage points to 93.17%,and the F₁ value increased by 2.94 percentage points to 95.35%.?2?In order to make the SSD real-time detection algorithm run on some embedded or mobile platforms with limited resources,replaced the basic network VGG in SSD with MobileNet V2 to reduce the amount of network model parameters and speed up the model running speed;added the CBAM attention module to the convolution module of MobileNet V2 to improve the expression ability of the network model;used deep separable convolution to replace the standard convolution in CBAM's spatial attention module to reduce the amount of network model parameters;used h-swish and h-sigmoid activation functions to optimize the ReLU and Sigmoid activation functions of the network model respectively,so that it can run better in devices with limited resources;optimized the network structure so that the network model can run better;used CutMix data enhancement strategy during the network model training process,in order to make full use of the image information in the training data set and enhance the network model training effect.Using the NREC's apple orchard dataset for experiments,the results showed that the improved MobileNet V2-SSD algorithm improved the accuracy of SSD detection by 2.04 percentage points to 98.31%,the recall rate increased by3.57 percentage points to 91.65%,the F₁ value increased by 2.87 percentage points to 94.86%,and the detection speed increased by 59.8%to 67.36 FPS.