Research On Target Detection Method Based On Convolutional Neural Network

Target detection is one of the key technologies of computer vision.Pedestrians and vehicles are hotspots and difficulties.This thesis studies pedestrians and vehicles as the target.The target detection method based on the sliding window is difficult to achieve real-time detection of the target due to the large amount of calculation.Although the convolutional realization of the sliding window object detection method solves this problem,since the detection bounding box is of a fixed size,the detection is inaccurate or even impossible to detect.In this thesis,by increasing the dimension of the output feature map,the network can accurately predict the position and size of the bounding box to achieve the detection of any size target.During the training process,the network constitutes the objective function with the error of the position and size of the bounding box,the category probability error,and the extent to which the bounding box encloses the target object.And by setting a detection target or detection area for each bounding box in advance,the convergence speed of the network is improved.In the prediction process,the optimal detection bounding box is selected by the category probability of the bounding box and the surrounding target object degree index,and the final detection result is obtained.By analyzing the problems in the target detection method based on the bounding box of arbitrary size,a full convolutional network target detection method based on prior knowledge is adopted.For detecting multiple or multiple types of overlapping targets,increase the number of detection frames and each detection frame has a set of category probabilities,and then remove the full connection layer to turn the network into a full volume containing only the convolutional layer and the aggregation layer.The neural network structure solves the problem of increasing the amount of parameters caused by detecting overlapping targets.For the detection of small target objects,the feature pyramid structure that fuses different size feature maps is adopted;and the network is avoided by introducing a priori size bounding box.When training,it falls into local optimum and improves the convergence speed of network learning.Finally,by integrating and processing KITTI and Udacity Self-Driving public data sets as training data sets,the appropriate subject network structure,training optimization method and hyperparameters are selected,and two methods are used to establish the target detection model.Through the evaluation,comparison and analysis of the model,it is verified that the full convolutional network target detection model based on prior knowledge is superior to the arbitrary size bounding box in detecting overlapping targets,small target object detection,model convergence speed and effect.The target detection model has achieved good detection results.