Research On The Application Of Machine Vision In Forklift Object Detection

With the development of domestic industrial technology and the increase in social logistics demand,the logistics and warehousing industry is booming,and Internet logistics,supply chain scheduling and smart logistics have become the industry’s vane.As one of the core links in smart logistics,warehousing has become the only way for the transformation and upgrading of the logistics industry.Most traditional warehousing in China not only occupies a large area and high construction cost,but also lacks intelligent development and reasonable planning,which makes accurate assessment of the energy efficiency level of warehousing an urgent problem to be solved.The existing research on the detection of forklift pallets in smart factories mainly includes three detection methods based on vision detection,various sensors or the fusion of the two.However,the data generated by laser sensors on the market is sparse and expensive,while traditional light sensors and weighing sensors are cheaper and more suitable for practical applications,but there are problems of low stability and low accuracy in practical applications.Based on the above problems,this paper is based on the deep learning of vision to recognize forklift pallet goods,and proposes an object detection algorithm in complex industrial scenes,so as to realize the intelligent and efficient management of the warehouse.The main tasks include:(1)Aiming at the problem of the lack of data on storage environment types in public data sets.We use industrial cameras to collect and expand data in the actual real storage environment,and then build and classify data sets of different storage tray states in different complex scenarios.According to the forklift operating time,the application scenarios are divided into two types: static scenarios with lower requirements for light robustness and dynamic scenarios with higher requirements for loading and unloading goods and inspection environment.(2)Aiming at the problem of small difference between adjacent pictures of self-built data set and huge amount of data,the image preprocessing operation of the on-site storage video is carried out,and the prepared training data set is selected to eliminate the information of the irrelevant target object in the image in the data set,and enhance the obvious regional difference The feature information of,to simplify similar data to the greatest extent,and finally to select training data with high image quality in a targeted manner.We first perform frame processing on the collected video and convert it into an image format,then measure the similarity of the images,compare the differences between two adjacent frames of images,and finally set category variables for clustering to ensure that images are selected in different clusters The biggest difference,effectively improving the performance of the training model.(3)For the above two types of scenes,considering the fact that the actual application is biased toward mobile device operations and different working hours,the model trains the feature extraction network for classification and detection,so as to obtain deeper hidden image information.Experiments show that this method has high detection accuracy and high illumination robustness,but the algorithm has low performance and features are too redundant.Therefore,improve the original feature extraction algorithm,focus on the lightweight operation of the network model,use a small amount of original convolution operation for output,and continue to use a series of simple linear changes to the upper network to generate ghost feature maps simply and efficiently.At the same time,the attention mechanism and multi-scale detection algorithm are introduced to extract key deep-level feature information in the image and improve the accuracy of the algorithm.Experiments show that this method compresses the trained model,uses multi-scale feature extraction to fuse image feature information,significantly improves the anti-interference ability of model detection,and tests and analyzes the improved algorithm in a self-built data set as a result.