Research On Object Detection And Model Acceleration In Open-Pit Mine

Although the mining industry has provided strong support for the development of manufacturing industry,the mining industry still faces the problems of low informatization and low operation efficiency.With the deep implementation of China’s manufacturing power and sustainable development strategy,the intelligence of China’s mining industry is continuously upgrading.For example,in order to solve the problems of recruitment and shortage of workers in mining factories,the mining industry is transforming to unmanned mode.However,most commercially deployed unmanned mining systems use Li DAR and Radar as sensors,but such sensors are difficult to accurately identify the type of obstacles,in particular,the type of distant obstacles,which will not be conducive to subsequent decision-making,but also affect the safety and overall efficiency of the unmanned system.In order to solve the above problems better,this paper takes open-pit mine target detection as the research object,analyzes the particularity of open-pit mine,and based on the existing detection algorithm,an innovative algorithm based on Yolov5 for obstacle detection in lightweight open-pit mines is proposed.The main advantage of the algorithm is that it can take the speed and accuracy into account,thus greatly improving the detection performance of different scale targets.The main work of this paper is as follows:Firstly,the network architecture and loss function of Yolov5 are improved as follows,which can achieve high precision detection:(1)in order to solve the problem of unbalanced feature utilization and lost features in the down-sampling method,based on the down-sampling method of existing deep learning target detection algorithm,a new down-sampling method based on bilateral feature with channel attention mechanism is proposed in this paper,which can improve the precision of feature sampling,especially for small targets.(2)For the purpose of enhancing the fusion effect of multi-scale features in Neck,channel shuffling and group convolution are proposed to replace the original channel concatenation in this paper,thus making the same object information of different scale features can be exchanged and fused in a group.(3)A more refined decoupled Head prediction branch is proposed,the three tasks of location,so that the three tasks of localization,foreground-background classification and inter-category classification can only focus on their own tasks,thus achieving effective prediction of each task.(4)The loss function,coupled location and classification tasks are optimized,so as to achieve joint optimization of two tasks.Secondly,in view of the lack of open-pit mine datasets,this paper collects and annotates mine datasets in different scenarios.At the same time,in order to enable the improved highprecision model to be inferred in real time on the deployment device,this paper uses channel pruning and layer pruning to perform different fine-grained compression acceleration for the highprecision model.Additionally,this paper also uses Tensor RT to further optimize and accelerate the pruned model.Thirdly,based on the dataset collected in this paper,the proposed YOLOv5s-Enhanced model is trained and achieves an AP accuracy of 59%,6.4% higher than the maximum model of YOLOv5,and 86% fewer parameters.After pruning and Tensor RT acceleration,the high-precision model can also achieve a compression rate of 43% in the number of model parameters,hence bringing about a fast inference of 53 FPS on the deployed devices.