Research On Dangerous Object Detection Algorithm Of EOD Robot Based On Deep Learning

With the development of social economy and the advancement of science and technology,robots have begun to replace humans in industrial production and daily life to complete some dangerous tasks.EOD robots are an important research branch in the field of robotics,which can replace humans in dangerous environments such as explosive areas,deep seas,and the universe to ensure the safety of workers.However,the EOD robots currently in use mainly rely on the video images returned by the robot's computer vision system,and then professional EOD personnel can remotely control the EOD robot to complete the disposal of dangerous objects through the keyboard or remote control.To a large extent,efficiency will be affected by factors such as the quality of the returned video image,the proficiency of equipment by professional EOD personnel,and labor costs.Therefore,the intelligentization of EOD robots is the key to solving the above limitations,and allowing EOD robots to automatically and quickly detect and identify dangerous objects is the basis for achieving intelligent EOD robots.Therefore,the study of a dangerous object detection algorithm suitable for EOD robots is of great significance for the intelligent development of EOD robots.In recent years,deep learning technology has continued to develop,and when it is applied to target detection,it has demonstrated excellent performance with high detection accuracy and high speed.Therefore,this article has carried out research on the Dangerous Object Detection Algorithm of EOD Robot Based on Deep Learning.The main content of the work is as follows:(1)Currently,there is no publicly available hazardous material data set.Based on this current situation,this article analyzes the types and characteristics of common hazardous materials in EOD work scenarios,and constructs the hazards that include eight types of hazardous materials such as grenades and landmines.In addition,in order to facilitate the training,verification and testing of the model,the self-built data set is labeled and divided;in addition,the diversity of the samples contained in the data set is very important for improving the generalization performance of the deep learning model Therefore,this article uses enhancement strategies such as translation,mirror flip,color change,and occlusion to expand the data set.At the same time,in order to build the most suitable algorithm model for detecting dangerous objects,the Faster RCNN,SSD and YOLOV4 are fully analyzed.The basic principles of the three target detection algorithms and their respective advantages and disadvantages,combined with the application scenarios of the EOD robot,finally selected YOLOV4 as the basic detection network for this experiment.(2)Aiming at the problem of low accuracy of the current dangerous object detection algorithm,this paper improves the model structure of YOLOV4 network,adopts the adaptive spatial feature fusion strategy(ASFF)to improve the quality of the extracted features of the model,thereby improving the detection accuracy of the algorithm;The K-Means clustering algorithm obtains the anchor point(Anchor)size that meets this experiment and improves the detection accuracy of the model;finally analyzes the composition of the data set.Because the data set constructed in this article has data imbalance,this article constructs a category balance The loss function improves the accuracy and generalization performance of the model;(3)Aiming at the problem of low efficiency of the dangerous object detection algorithm,this experiment uses Mobile Net V3 as the backbone feature extraction network of the model,which effectively reduces the number of model parameters and the amount of convolution calculations,and improves the efficiency of model detection.