Research On Key Technologies Of Object Detection Based On Improved SSD Network

With the development of artificial intelligence technology,smart driving vehicles have become a research hotspot in the automotive industry.Object detection is a key technology for smart driving car's environment awareness.Accurate detection of forward traffic objects can help make correct driving decisions.Object detection of video images based on machine learning is one of the main means of smart driving car's environment perception.Compared with traditional machine learning detection algorithms,the deep learning detection algorithm represented by SSD(Single Shot MultiBox Detector)has obvious advantages in image feature extraction and running speed,and can obtain better detection performance,but the detection accuracy cannot yet fully meet the need of driving scenes.Therefore,it is of great theoretical and practical significance to study the object detection method based on improved SSD network to improve its detection accuracy.Based on an in-depth analysis of the characteristics of SSD network,this paper uses the deep residual module and multi-layer feature fusion structure to improve the deficiencies of detection framework.For the unbalanced sample in training,the introduction of the modulation parameter with parameters in the loss function is introduced,and the issue can be solved.The main work and contributions of this paper are as follows:(1)This paper first analyzes the basic principles of the SSD detection algorithm,and then according to the detection tasks in the driving scenario,select the appropriate features to extract the backbone network,adjusting the prior box mechanism in the SSD algorithm based on the statistical data,and then adjust the multi-task loss function.Finally,a framework for detecting network is established.(2)For the problem of unbalanced samples in the SSD model training process,based on the principle of hard sample mining,this paper introduces the modulation items with parameters,improves the original cross-entropy loss function design.Deductions of the forward and backward propagation process of the loss layer have been made.The method can make the improved loss function automatically increase the weight of the loss value of the hard sample during training,so that the model can be trained more fully and the convergence speed of the model can be accelerated.(3)The SSD network uses multi-layer feature maps for object prediction.The shallower feature layers need to obtain low-level features and high-level semantic features at the same time,which increases the difficulty of network model training and learning and is not conducive to the improvement of detection accuracy.To solve the problem,this paper proposes a prediction supplement module based on the deep residual structure.This method can separate the prediction layer and the feature layer structurally,so that the feature information of multiple layers to be detected is more abstract,which can improve the detection accuracy of the model.(4)Because the semantic information of the low-level feature map is relatively insufficient,the detection effect of the SSD network on the small object is poor.To solve this problem,this paper proposes a multi-layer feature fusion structure based on feature pyramids.The high-level feature maps are up-sampled in the form of deconvolution and then connected to lower-level feature maps.This method can effectively enhance the semantic information of low-level feature maps,so as to improve the quality of the entire network feature extraction,thereby improving the detection accuracy.The above improvements are combined to form a complete object detection network based on an improved SSD.The paper uses the KITTI data set to train and test the improved network model,and then conducts a number of comparative experiments.The experimental results show that the above improvements can improve the detection accuracy of the network in different degrees.At the same time,the detection effect of this network is also superior to the more advanced Faster R-CNN and DSSD networks,and the detection speed can reach the real-time level,which can meet the actual application requirements.