Research On Real-time Detection Technology Of Deformable Construction Machinery

With the acceleration of urbanization,land resources have become scarce,and economic development has brought about problems of illegal land occupation and illegal construction in recent years.In view of the reality that land violations are difficult to find,this paper is devoted to the research of real-time detection technology of deformable construction machinery when the patrol vehicle moves.The main research contents are as follows:(1)A fast annotation algorithm for deep learning image dataset based on HOG features is proposed.The algorithm obtains the position of the targets based on the Gaussian mixture model,and automatically gets the targets' categories based on the Mean Shift clustering of HOG features.After the above steps,a weak classifier of digger is acquired.The labeled digger samples obtained is used to train the Deformable Part Model to annotate more unlabeled images,moreover,artificial fine tuning labeled data software is developed to correct partial inaccurate labeled samples' position.Experimental results show that this algorithm can quickly complete the automatic annotation of image datasets,manual annotation of 1000 images takes about 60 minutes,while our fast annotation algorithm can complete the same work within 10 minutes,therefore,this algorithm can reduce the cost of acquiring labeled datasets and greatly improve the effectiveness of acquiring labeled datasets in deep learning.(2)An improved multi-view SSD based on jump fusion of feature maps is proposed.On account that SSD has much worse performance on smaller objects than bigger objects,a novel feature fusion method which means jump connection between multi-scale feature maps for detection is proposed to improve the detector's sensitivity and impact of small objects,at the same time,it helps improve the overall detection performance of SSD.To solve the problem that the model's detection accuracy of the deformable construction machinery decreases due to the rapid change of viewing angles,the multi-view strategy is integrated into the model to have a good performance.The experimental results show that the improved algorithm has achieved good results on the detection of small objects and the detection of diggers in high dynamic scene,the mAP of digger has increased from 90.6% to 96.9%.(3)A detection system combined with visual tracking is built for real-time detection of deformable construction machinery.With SKIPSSD as the detection model and KCF tracker as the tracking model,a stable and reliable digger real-time detection and tracking system is built.The experiment proves that our system's time of detection only takes up 1/3 or less of the SKIPSSD model's,and the higher the resolution of images,the more sharply detection speed increases.Our system utilizes SKIPSSD to make up for the defect that the KCF algorithm can not retrack the target after it is lost.The continuity and the high real-time performance of visual tracking algorithm are used to reduce the leak detection rate and improve the speed of detector.