Detection And Implementation Of Occlusion Target Based On Depth Neural Network

In recent years,artificial intelligence is become more and more popular in today's society.The concept of artificial intelligence permeates all of life.Because machine vision is an indispensable technical aspect to realize artificial intelligence,the importance of image application is self-evident.Computer interprets image content through object detection technology.Because occlusion problem is a common and inevitable phenomenon in real life,so the detection of occluded object is not only an important branch of object detection,but also a significant subject to be solved.This paper is based on SSD target detection algorithm.This paper is aimed at the problem of dense same kind of targets occluding each other and targets occluded by other interference items in the complex environment background.Taking pedestrian target and vehicle target as examples,this paper makes a profound study.The specific results are as follows:1.An algorithm of occlusion detection based on improved NMS(Non-maximum suppression)is proposed NMS is a redundant prediction frame with large overlap in target detection,and the determination of the final residual prediction frame is the key step of the real target location.When traditional NMS processes two very similar targets,the frame with lower score will be directly suppressed because the intersection ratio with the highest score frame is greater than the preset threshold value,which leads to the problem of missing detection.On the basis of adding a new threshold value in soft NMS to punish the weight,the second threshold value is further preset to delete the frame to be detected,which is too large in the ratio of partial intersection and union,so as to reduce the probability of repeated detection of the same target.By setting the third threshold,the parallel NMS algorithm can directly separate the dense targets from each other to reduce the detection time.The experimental results show that the algorithm can solve the problem of missed detection,false detection and repeated detection of similar dense occluded objects.2.An algorithm of occlusion detection based on enhanced Repulsion Loss is proposed.In this paper,the original loss function of SSD algorithm is replaced by the loss function of Repulsion Loss.By combining an advantage function of reducing the distance between attractive items and increasing the distance between two rejection items,Repulsion Loss can enhance the prediction ability of SSD algorithm to prevent the missing detection of occluded objects.Because of its application in dense pedestrian detection,through the analysis of its idea and experiment to explore the universality of its application in other dense target detection.Taking the vehicle target as the object to verify the idea,extend the application of general occlusion target detection.This paper analyzes the shortcomings of Repulsion Loss,and adjusts the proportion of rejection items flexibly and scientifically to enhance the function of Repulsion Loss.On the other hand,starting from the feature extraction features of SSD network,this paper focuses on six common convolution layers of six feature graphs generated in SSD network.This paper uses the structure of Inception-ResNet-v2 network to change its feature generation mode,which makes the feature graphs contain more information,This information includes more feature information about occluded objects and better detects occluded targets to distinguish other jamming targets.In addition,the improved algorithm model is trained by using the classical data set.Then,the self built data set is added to form the mixed data used for experiment again.The experimental results show that the improved algorithm has a good accuracy for dense occlusion detection and complex background occlusion detection.The generalization ability of the model has been further improved.3.The design and implementation of target detection software are carried out.This software is based on Python language.The interface of software is designed by using pyqt5.The secondary functions of software is designed by using OpenCV visual library.The trained model is embedded into the software to realize the functions of pedestrian and vehicle target detection in road traffic image.