Research On Real-Time Pedestrian Detection Based On Convolutional Neural Network

In recent years,with autonomous driving becoming a hot item in global research,pedestrian detection technology in the case of road traffic is one of the core tasks of autonomous driving technology,and it is receiving more and more attention.The pedestrian detection problem in the traffic environment has high requirements for the detection accuracy and detection speed of the algorithm.Although the traditional pedestrian detection method can meet the speed requirement,the difference in detection accuracy is far.The target detection algorithm based on convolutional neural network has excellent detection performance,but the performance of the target detection algorithm is not directly applied to the pedestrian detection problem.good.Aiming at this problem,this paper mainly constructs a pedestrian detection method with higher precision and faster speed based on the target detection algorithm.In order to solve the problem that the target detection algorithm performs poorly on the pedestrian detection problem,this paper mainly does the following four aspects:(1)Based on the leading two-stage detection framework Faster R-CNN in the target detection field,combined with pedestrians Single size information,improve the size of the network anchor box,and then experiment with the Caltech pedestrian data set on the deep learning framework Keras.The experimental results show that the algorithm improves the pedestrian detection accuracy,but its detection speed is still very slow;(2)in the single Based on the SSD of the stage detection framework,the experiment was carried out with the same improved method as Faster R-CNN.The experimental results show that the algorithm improves the pedestrian detection speed,but it can not maintain the same detection accuracy as Faster R-CNN.(3)Compare two Experiments have analyzed two reasons that affect the accuracy and speed of pedestrian detection.First,the design of candidate region generation network(RPN)in Faster R-CNN is an important factor to ensure its detection accuracy,and the second is the single-stage detection method of SSD.Is an important reason to improve its detection speed;(4)for the construction of higher precision and faster pedestrian detection method,targeting For the two reasons,this paper has made the following three improvements on the basis of the single-stage detection algorithm SSD:First,the resnet50 truncated network with the best image classification effect is used as the backbone network of the SSD algorithm to extract image features.Second,to adapt to pedestrian detection,increase pedestrian size information and adjust the size of the network anchor box;third,improve positioning and increase the IOU threshold..Referring to the two-step prediction mechanism of default anchor box in Faster R-CNN,one step is RPN prediction,the other step is ROI prediction,and RPN prediction is applied in SSD algorithm.By stacking a series of predictors,the SSD is gradually evolved directly.The default anchor box is used to improve the positioning,and then use the IOU threshold multiple times to mine the more difficult negative samples,thereby improving the detection effect;Most importantly,a new single-stage pedestrian detector has been proposed that not only enjoys the speed of the SSD but also maintains the detection accuracy of Faster R-CNN,and achieves a good detection effect on the Caltech pedestrian detection benchmark.