Study On Multi-target Detection Based On Deep Convolutional Neural Network And Image Sensor

In intelligent transportation,robot research,pilotless driving and visual surveillance area,researching on detecting of multiple kinds of targets on roads has great academic significance and engineering significance.The environment of traffic is complex.Meanwhile,the target feature extraction is greatly influenced by weather,shelter and other factors.Therefore,the detection,classification and tracking of multiple targets on the road is a comprehensive problem with multiple challenges.The existing target identification framework is mainly divided into two types.One is based on RPN(Region Proposal Network)represented by Faster R-CNN.The recognition accuracy is high,but the recognition speed is very low.The second one is based on regression,which uses the idea of regression and is represented by the YOLO network.The recognition accuracy is low,but the processing speed is very fast.At the same time,all of them are based on the open data set which is not powerful for objects in traffic field.Considering the traffic environment in China,the problem of incomplete and inaccurate still exist.In view of the above problems,this paper improves the Faster R-CNN and YOLO algorithms and builds 2 new networks.First of all,this paper expands the public datasets VOC2007 and VOC2012,enhances the original road targets through web crawling and collecting photps,and adds new categories such as signal lights.Afterwards,a simple evaluation of the popular deep learning framework was conducted,and TensorFlow was chosen as the experimental framework.Use the enhanced data set to train the improved GoogLeNet.The improved approach is to add the spatial pyramid pooling algorithm,use the depth separation convolution module to improve the Inception module,and add the batch normalization layer.This allows the network to use images of different sizes as input and accelerate training.The process converts low-level images to high-level semantics,and then extracts the features of a single static image.This improvement improves the accuracy and speed of feature extraction.After that,this network was used as a feature to extract the network,and a road-based multi-target detection network based on regional selection and a road-based multi-target detection network based on regression were constructed.Road-based multi-target detection network based on Faster R-CNN,the classification network obtained in the previous chapter was used as a feature extraction network,and detection was performed using ROI pooling.Through the experiment,the anchor selection strategy,learning rate and batch size are adjusted to obtain the best training effect.The experiment proved that the network mAP increased by 2.5%and the detection speed of the standard input increased by 0.1s.and for actual input images the detection speed increased by about 9s.The regression-based road multi-objective detection network is based on YOLO and uses the classification network obtained in the previous chapter as a feature extraction network to select a candidate frame shape that is most suitable for road targets through a K-means clustering algorithm.Through the experiment to complete the parameter tuning.Under the data set of 2,the mAP increased by 1.9%and the actual input detection speed increased by approximately 0.3s.Finally,the training experience in road target detection is summarized,and the two network models are compared and analyzed.Experiments show that both can complete the detection of the main objective of road traffic in the real environment.The multi-target detection network based on regional selection has higher accuracy(60.3%)and slower speed(3.7fps),and the regression-based road multi-target detection network has lower accuracy(41.1%)and faster speed(34fps).Therefore,it needs to be flexibly applied in combination with different usage scenarios.