Research And Application Of Object Detection Algorithm Based On MobileNetV2

With the development of deep learning and the improvement of hardware,target detection has gradually become one of the important research contents in the direction of computer vision.It has automatic and intelligent attributes,can determine whether the image contains the specified target,and give the accurate location information of the target,which is widely used in intelligent transportation,video monitoring and many other scenes.However,in the case of complex image background and target occlusion,it is difficult to achieve better detection effect.Moreover,the use of convolutional neural network improves the complexity of the model,which leads to the increase of calculation cost,resulting in problems such as slow calculation speed when the algorithm is applied to equipment with low calculation force.Based on this,this paper studies the target detection method based on deep learning deeply,constructs a lightweight target detection model,and applies the model designed in this paper on the self-made armor plate data set.Among them,the main work is as follows.1.In the task of target detection,the traditional methods mostly use the manually designed feature algorithm to classify the image,so there are some problems such as poor robustness and long time consuming.Based on this,this paper introduces convolutional neural network as the dependency structure of target detection.Based on the analysis of the structure of convolutional neural network,the detection models based on single-stage and two-stage are studied respectively.The anchor frame regression method of single-stage detector(SSD)is used to design the network model,and the prediction of the target is realized on the characteristic map of different scales.2.Aiming at the problem of large parameters and low detection efficiency of the existing SSD model,the deep separable convolution and inverted residual block structure in mobilenetv2 network are introduced as the feature extraction network of the model,and the multiscale target detection model with mobilenetv2 network as the backbone network is designed to realize the pruning and lightweight algorithm model of the network;In view of the problem that NMS operation in post algorithm is easy to make similar targets with high overlap miss detection,this paper introduces the confidence decay function and puts forward an improved non maximum suppression strategy.Through tensorflow framework,the model is designed,and the improved algorithm is carried out on VOC data set with mean average accuracy,parameter quantity and frame rate as performance indicators Performance verification,experiments show that the proposed target detection algorithm based on mobilenetv2 and improved NMS has good performance in subjective evaluation and objective evaluation.3.The improved mobilenetv2-ssd model is applied to the robomaster projectof the national university robot competition to realize the detection of the armor plate on the intelligent vehicle.The self-made armor plate data set is constructed,and the model is trained through parameter adjustment.The armor plate detection experiments under different environments are designed,which achieve the effect of fast and accurate identification of the target to be tested,and finally achieve 92.4% AP performance index.The experimental results show that the improved detection model has higher accuracy and faster detection speed.It can be used in low calculation force equipment to detect armor plate in complex environment,and the practical application effect is better.