Research And Application Of Object Detection Algorithm Based On SSD

As the most basic part of computer vision,object detection is one of the research difficulties and hotspots in this field.In recent years,with the rapid development of deep learning,the Single Shot Detector(SSD)algorithm based on regression,that is,the one-stage object detection algorithm which directly predict the coordinates of bounding box and categories,has the advantages of real-time performance and high accuracy,and therefore gets a lot of attention.However,in practical engineering applications,there are still some problems in the existing SSD algorithm,such as low detection accuracy for small objects,partially occluded objects and low-resolution objects,as well as large number of model parameters and computational complexity and thus inefficient operation on low-cost GPU hardware platforms or embedded chips.In order to solve these problems,the object detection algorithm with higher detection accuracy and faster detection speed is studied and applied to the object detection of actual scene.It can not only meet the needs of practical engineering application,but also promote the large-scale use of SSD algorithm,and accelerate the development of object detection technology,which has important theoretical and practical significance.In this paper,the basic theory of the SSD algorithm is deeply studied,and the method of fusion of deep and shallow features with context semantic information is used to improve the detection precision of small objects;the method of coarse to fine in which the regression of anchors is decomposed into two steps is used to improve the detection accuracy of partially occluded objects;the method that replacing some convolution layers of backbone network and multi-scale feature extraction network with receptive field block is used to improve the capability of feature extraction and the accuracy of low-resolution object detection;meanwhile,the lightweight method that replacing some standard convolution with depth-wise separable convolution is used to improve the detection efficiency.Furthermore,several new object detection algorithms are proposed and deployed on low-cost hardware platforms and embedded chips,meeting the requirements of object detection in different scenarios,such as universal object detection,pedestrian and oxygenator detection under fishery monitoring,vehicle marker detection and excavator detection under monitoring scenes.The main research contents and innovative works of this paper include the following aspects:(1)The detection performance of SSD object detection algorithm under low-cost GPU is analyzed in detail.Firstly,the key techniques,training skills and evaluation indexes are analyzed,and then SSD is tested on the common data set PASCAL VOC and low-cost GPU hardware platform and embedded chip.The experimental results show that the existing SSD algorithm fails to detect small objects,occluded objects and low-resolution objects,and the detection efficiency of SSD on low-cost hardware platform and embedded chip is low.The main reason is that the ability of feature extraction is not enough,and the model size and the computational complexity is too large,which indicates the direction for the improvement or optimization of the SSD algorithm.(2)A fast object detection algorithm based on feature fusion is proposed.Aiming at the problem that the model size and calculation amount of the existing SSD algorithm is too large to efficient operation on low-cost hardware platform,Mobile Net V2 is introduced and the standard convolution is replaced by the depth-wise separable convolution to keep a lightweight model size.Then,the Depthwise-Feature Pyramid Network(DFPN)is combined to redirect the information flow from deeper and smaller feature maps to shallower feature maps,and the semantic information feature maps are fused to enhance the context information and improve the feature expression ability.Finally,the proposed algorithm,Mobile Net V2 SSD and Tiny-DSOD are compared and tested on the PASCAL VOC test set on the low-cost hardware platform and the embedded chip.The experimental results show that the m AP of our algorithm is 74.51%,which is 4.46% and 2.55% higher than Mobile Net V2 SSD and Tiny-DSOD.At the same time,the detection speed of our algorithm is 11.8 ms,which is basically the same as Tiny-DSOD,showing that our algorithm has better precision and maintains real-time detection.(3)An object detection algorithm based on coarse to fine is proposed and applied to practical detection scenes.The method from coarse to fine is further studied and introduced into SSD.The anchor boxes are selected by using coarse to fine prediction mechanism and the fusedAnchor Refinement Module(fused-ARM)module fusing context feature maps is proposed to give a better result of selecting anchor boxes.Then,the Transfer Connection Block(TCB)feature fusion module is utilized to fuse the deeper and shallower feature maps and the Feature Re-extraction Module(FRM)is added to enhance the feature extraction ability.Finally,the comparison test is implemented on PASCAL VOC test set and low cost hardware platform,showing that the m AP of the proposed algorithm is 80.7% and the FPS is 27.7fps and 37.9fps.Meanwhile,its application on pedestrian and oxygenator detection in fishery monitoring scene arrives the m AP of 90.1% and 90.6%,respectively.In the detection area of the oxygenator,the Harris algorithm is used to detect water flower corner points and the Lucas-kanade optical flow method is used to calculate the optical flow rate,and the average water flower corner point displacement of each frame is obtained.The oxygenator working state detection is realized by the Support Vector Machine(SVM)algorithm.The average accuracy of the proposed method is 99.78%,showing that the proposed algorithm can be applied to detect the oxygenator working state under different illumination,angle and distance conditions with strong robustness and real-time performance.(4)An object detection algorithm based on receptive field block is proposed and applied to practical detection scenes.The parameters of the anchors are optimized according to the relationship between the anchor boxes and the receptive field.Then the adjacent shallow feature map of RFBNet is fused by the method of context fusion.Finally,the proposed algorithm and some baseline algorithms are compared and tested on the PASCAL VOC test set.The experimental results show that the m AP of our algorithm is 80.49%,which is 1.8% higher than the RFBNet and 3.6% higher than SSD.At the same time,the m AP of our algorithm on our vehicle markers is 94.03%,which is 1.87% and 10.65% higher than RFBNet and SSD,respectively.The results show that the proposed algorithm has better detection accuracy and robustness to low-resolution objects.(5)A lightweight object detection algorithm for embedded chips is proposed and applied to practical detection scenes.Using the Depthwise Dense Block(DDB)to form the backbone network,the expression ability of the feature is improved,and the multi-scale feature extraction module Bottleneck Down-sampling Module(BDM)is designed.Then,the application of the proposed algorithm on the PASCAL VOC test set arrives the m AP of 69.5%,and the time of 145.2 ms on the embedded chip.Compared with the Mobile Net V1 SSD and Tiny-DSOD,the time was 70.8 ms and 11 ms less,respectively.At the same time,the proposed algorithm is applied to the excavator detection under monitoring scene,achieving the m AP of 90.6%,and then in the detection area of the excavator,a mixed Local Binary Features(LBF)shape regression model is established for various postures of the working device of the same brand excavator to predict the shape information of the working device of the excavator,and the working state characteristic description of the excavator is constructed.Finally,the working state detection of the excavator is realized by using the SVM classifier.The accuracy rate of the proposed method is 93.53%,showing that the method has good detection precision,and the influence of the change of the shape caused by the illumination and the multi-attitude is overcome.In conclusion,through the work above,it turns out that on the basis of SSD,using the method of feature fusion,coarse to fine and receptive field block could improve the accuracy of object detection,and solve the problem of poor performance on small objects,partial occlusion objects,and low-resolution objects.Meanwhile,combining the lightweight method with SSD could reduce the number of model parameters and the calculation amount and enhance the operation efficiency on low cost GPU hardware and embedded chips.Therefore,based on the SSD algorithm,this paper proposes algorithms with high detection accuracy while keep comparative real-time performance and broadens their application scene,which is a very frontier research direction.