Research On Pedestrian Detection Technology Based On Deep Learning

With the rapid development of deep learning technology in the field of computer vision,pedestrian detection technology has made good progress,but there are still some problems in some harsh environments,such as dense pedestrian detection and pedestrian detection with different scales.The development of pedestrian detection technology plays a great role in promoting intelligent monitoring,interactive robot,pedestrian re-recognition and other related technologies.Therefore,it is extremely important to solve the technical problems of pedestrian detection.In view of the above problems,do further research,the main innovation points of this paper are as follows.1)For pedestrians size problem,this paper modified YOLOv3 anchor box parameters in algorithm,and multi-scale feature fusion module is introduced in feature extraction Darknet53 internal network,changing the original way of feature extraction,feature extraction network depth up to improve the ability of models of nonlinear said,further enhancing scales pedestrian feature extraction ability.The experiments show that the average accuracy of the improved algorithm is 5.49% and2.26% higher than that of the benchmark algorithm after the training in Caltech and ON?MERGE data sets.2)For some dense occlusion problems,SE-RES?UNIT with attention mechanism was used in this paper to replace the YOLOv3 model of RES?UNIT in Darknet53.Under the guidance of the attention mechanism,the model can pay more attention to the visible features of pedestrians,reducing the response of the features of the blocked part of pedestrians to the channel,improving the representation of important information by the model,and reducing the interference of background information to the model.The experiments show that the average accuracy of the algorithm using SE-res?unit is 5.87%,5.08% and 3.77% higher than that of the benchmark algorithm after the training of Person dataset in MS COCO,ON?MERGE and Caltech respectively.3)In order to improve the generalization ability of the model,this paper firstly through data enhancement technology for different scales of data set image stretching,rotating have different attitude,different shade degree of pedestrian,fusion of different concentration mechanism in feature extraction network,and model can be more different attitude,different shade degree of pedestrians.The experiments show that the average accuracy of the algorithm combining channel attention,spatial attention and mixed attention is 3.3%,4.97% and 9.58% higher than that of the baseline algorithm after the training in Caltech data set.The training results in ON?MERGE data set are 1.83%,4.18% and 4.31% higher than the baseline algorithm respectively.4)For difficult cases,a detection model based on cascading strategy is designed in this paper.After many experiments,SE-RES?UNIT replaced the attention model of RES?UNIT in Darknet53 with the best detection performance.Therefore,both the first stage model and the second stage model were set as SE-RES?UNIT attention model.The detection of difficult examples of the first-stage model can be used as the training input of the second-stage model,and the second-stage model can have stronger ability to identify difficult examples.The experiment shows that the cascade model increases the overall recall rate by 1.71% on the premise of not reducing the detection accuracy of the first-stage model.