Pedestrian And Multi-target Detection Based On Deep Learning

Target detection technology is a key research topic in the field of photoelectric detection imaging and computer vision.Traditional target detection methods are difficult to adapt to diverse scenes,and the detection effect is not satisfactory as well.Nowadays,Convolutional Neural Network(CNN)has made major breakthroughs in image classification and recognition,bringing target detection technology to a new level.Based on the fully study of CNN,targeting on the problems caused by current network model such as low accuracy rate and low target recall rate,this paper proposes a GoogLeNet convolutional neural network model with fusion reduction layer for pedestrian detection.Meanwhile,based on this,use convolutional neural network model based on SSD(Single Shot Multibox Detector)for pedestrian and multi-target detection.The main research contents and innovations of this paper include:(1)A GoogLeNet convolutional neural network pedestrian detection system combine a 1*1 convolutional layer was designed.The Inception model is replaced by the 3x3 and 5x5 convolutional layers followed 1*1 convolutional layer,which could effectively reduce the quantity of parameters,improve the convergence of the network,and enhance the expression capability of the model's feature.(2)Optimizing the positioning frame by using the loss function.The loss function is redefined according to the improved model.The output shows that as the number of network iterations increases,the loss gradually fits the noise and the non-representative features in the training dataset.The improved model was tested on the VOC2007+2012 dataset.The result shows that the pedestrian detection model achieved a precision rate of 93.1%(1.6% better than the AlexNet model)and achieved an AUC value of 0.96 on the ROC curve(0.03 better than the AlexNet model).In the real data evaluation,the pedestrian detection model achieved 95.47% of precision rate(4.11% better than the AlexNet model).(3)Targeting on the low target recall rate problems of GoogLeNet convolutional neural network model,by using convolutional neural network model based on SSD(Single Shot Multibox Detector)for pedestrian and multi-target detection,and by evaluating model performance with real data.The results show that for the multi-target detection task of pedestrians,bicycles and cars,the model achieved a mean precision rate of 89.47%(11.6% better than the Faster R-CNN model,16.81% better than the YOLO model).For the pedestrian detection task,the small target pedestrian recall rate is 82.97% with this model(18.02% better than the AlexNet model,20.04% better than the improved GoogLeNet model).