Research On Road Pedestrian Detection Method Based On Faster RCNN

With the development of artificial intelligence,deep learning methods are gradually applied to the intelligent vehicle assisted driving module,such as pedestrian detection.At present,the difficulty of pedestrian detection is mainly the problem of pedestrian multi-scale and pedestrian occlusion.In the final analysis,it is to improve the detection speed and accuracy of the model.The paper uses Faster RCNN algorithm to improve the speed and accuracy of pedestrian detection.The main research contents are as follows:（1）Aiming at the multi-scale problem of road pedestrians,this paper improves the original RPN network structure.First,nine different scale anchors were used to extract candidate regions for multi-scale pedestrians,and then use 1×1,3×3,5×5 sliding windows on the feature map of the last layer of the VGG16 network to generate candidate regions to increase the robustness of multi-scale pedestrian detection.The original RPN network feature extraction time reached 0.06s/frame,and the total detection time reached 1.03h.Compared to the original RPN,the improved RPN saves 0.04s/frame in feature extraction time,saves 0.38h in total detection time,and improves the detection speed.（2）Aiming at the occlusion problem of road pedestrians,this paper adds feature fusion technology to the algorithm.Feature maps are generated at the Conv4₃ and Conv5₃ layers in VGG16,and the candidate regions are mapped onto these two feature maps to get the feature map after fusion.After ROI Pooling and L2 regularization,a fixed-size feature vector is obtained,and finally sent to the subsequent fully connected layer to implement pedestrian detection and range box regression.The detection rate of the model after feature fusion reached 81.9%,which was about 13.9%higher than the LatSVM model,and the accuracy of the entire model was enhanced.（3）Finally,the improved RPN network structure and feature fusion technology are cascaded,and the improved Faster RCNN model structure is obtained.Repeated experiments on different public data sets show that using the original Faster RCNN network model results in a detection speed of 0.12s/frame and an accuracy rate of 87.3%.The improved Faster RCNN network model has a detection speed of 0.04s/frame and an accuracy rate of 91.5%,which is increased by 0.08s/frame and 4.2%,respectively.Its detection speed and detection accuracy have reached the expected requirements.Finally,experimental verification was carried out in actual road scenarios,and the accuracy rate reached 89.9%,indicating that the algorithm has certain practical value.