Research On Pedestrian Detection Algorithm Of Convolutional Neural Network Based On Multi-Scale Fature Fusion

Pedestrian detection is a classical problem in computer vision,and has always been a research hotspot in the field of computer vision.Due to the particularity of its detection target,wide application prospect and commercial value,it has become the object of research by scholars and practitioners at home and abroad.In recent years.with the rapid development of deep learning,the pedestrian detection algorithm based on computer vision technology combined with deep learning method has gradually become the mainstream algorithm in the field of pedestrian detection.In particular,since the deep learning algorithm based on convolutional neural network was proposed,pedestrian detection algorithm has achieved a breakthrough development,which has achieved high detection accuracy and speed,making pedestrian detection efficiency reach a new levelHowever,based on a comprehensive literature analysis,the main challenge for pedestrian detection based on deep learning is the gap between the accuracy and speed of detection algorithm and the actual requirements,and its detection accuracy and speed still need to be further improved.Among the m,the tradeoff between pedestrian detection accuracy and speed is the most challenging problem,and multiple pedestrian multi-scale problems are one of the important influencing factors.In the design of pedestrian detection algorithm,the resolution of the multi-person multi-scale problem can improve the accuracy of pedestrian detection,so it is of practical significance to study the multi-scale problem of pedestrian detection.Therefore,in the face of such problems,on the basis of full investigation and careful analysis of current representative detection algorithms,a convolutional neural network pedestrian detection algorithm based on multi-scale feature fusion is proposed.In order to make full use of the various features of pedestrian image to realize pedestrian detection,this algorithm fuses the features of image such as color and texture to enrich the information of the feature image finally obtained,so as to effectively reduce the number of false detection and improve the detection accuracy.In addition,feature graphs of different scales were extracted from different locations for prediction.and the prediction results were finally combined.Compared with the prediction at the last position of the network,it can effectively reduce the loss rate and improve the detection accuracy.The main research work includes the following three aspects(1)Design pedestrian detection network structure.According to the experience gained by the recurrence algorithm,the classical target classification network structure is used for reference,and it is fine-tuned as the basic feature extraction network.In addition,a multi-scale feature fusion method is proposed,which USES the up-sampling and magnification feature map to fuse the features and fully extract the pedestrian information from various aspects,including color texture information and gradient information.The feature information learned is enriched,which makes up for the lack of comprehensive feature information extracted from single feature and detailed description of pedestrian features,so as to improve the accuracy of detection results.(2)An anchor frame setting method suitable for pedestrians is proposed.Due to the pedestrian has the characteristics of non-rigid deformation,and pedestrians attitude in the image of the diversity,when shooting Angle is different also,cause and pedestrian scale sizes in image or video,so set up different scales and different aspect ratio of the anchor box to forecast image of pedestrians,the make the final test box more accurately(3)The designed network model was trained,and the trained model was experimentally verified on different pedestrian detection test sets.Meanwhile,the experimental results were analyzed and compared quantitatively.The experimental results show that the algorithm proposed in this paper can effectively realize the task of pedestrian detection,and compared with other algorithms,the algorithm in this paper has a higher accuracy,and the detection speed does not decrease.The algorithm is proved to be feasible and effective.