Research On Container Stack Twist-Lock Detection Method Based On Industrial Computer Environment

With the rapid development of China’s economy and the booming import and export trade,the safety management of port containers arises at the same time.Stack twist-lock is an important component of containers,which is mainly used to connect and fix trucks with containers or containers with containers.In the course of transportation,due to the operator’s forgetting to dismantle the stack twist-lock,other containers connected by the stack twist-lock are often fell off from air during lifting.Consequently,it will cause accidents and economic losses.The research on using computer vision technology to detect and track container lock and warn in time before the accident happens has important theoretical value and practical significance.The main difficulty of stack twist-lock detection lies in the fact that the industrial computer with low storage performance is often used in practical terminal applications,and the existing deep learning methods are difficult to meet the real-time requirements.In addition,the size of the stack twist-lock itself is small,and the background of vehicles,personnel and ships in the terminal scene is complex,which increases the difficulty of stack twist-lock detection.In view of the above problems,this paper makes a systematic research on stack twist-lock detection in industrial computer environment from the two perspectives of adapting to complex outdoor environment of port and low performance of industrial computer.To solve the problem of complex outdoor scene of port,a fast stack twistlock detection method based on multi-model integrated is proposed,which improves the real-time performance of stack detection.To solve the problem of low storage performance of industrial computer,a convolutional neural network compression algorithm based on compressed sensing is proposed to reduce the space complexity of the stack twist-lock detection algorithm model.The related work is as follows:1.In view of the complex outdoor scene of port,this paper proposes a fast stack detection method based on multi-model integrated which is capable of adapting to industrial computers with low computing performance.The algorithm consists of three modules:in the container lock detection module,YOLO-v2 object detection algorithm is used to locate the location of the lock in the first frame of the video stream to obtain the initial search area of the stack lock;in the container motion tracking module,according to the initial position of the lock hole obtained by the previous module,this paper uses the improved KCF algorithm to continuously track the container motion,and updates the height information of the container in real time until the key frame area is reached.In the multi-model integrated classification module,this algorithm calculates the candidate area of the stack.Finally,a simplified VGG network structure along with an ensemble learning method is used to detect the stack twist-lock of candidate regions.2.Considering the low storage performance of industrial computer,this paper proposes a compression method of convolution neural network based on compressed sensing from the point of view of optimizing the storage performance of convolution neural network.The algorithm includes two modules:compression module and decompression module.The weights of convolutional neural networks are highly concentrated in the low frequency region in frequency domain.The weights of convolutional neural networks in frequency domain are pruned,and the sparse weights are obtained.Then the compressed weights are obtained by dimension reduction sampling of the sparse weights.In the decompression part,the proposed method uses the sparse vector restoration algorithm based on compressed sensing to restore the compressed weights to the sparse vectors in the frequency domain,and finally convert them to the original weights in the spatial domain.In the experimental part,the performance of each module of the stack twist-lock detection algorithm proposed in this paper is evaluated and compared with other methods.The results show that the performance of the proposed stack twist-lock detection algorithm is better than comparative methods.In addition,the proposed convolutional neural network compression algorithm is also compared with other convolutional neural network compression algorithms.Because the importance of each weight value of convolutional neural network is taken into account in pruning,the performance of the proposed algorithm is better than that of the comparative method when the pruning rate parameters are set higher.The convolutional neural network compression method proposed in this paper can be combined with the proposed stack twist-lock detection algorithm to achieve higher operation and storage efficiency on industrial computer with lower performance.