A logistics overlap identification method based on deepstream

Abstract

The present invention discloses a logistics overlap identification method based on DeepStream, belonging to the field of logistics sorting technology. The method includes target recognition neural network overlap identification, which includes simultaneously using two target detection neural networks to detect the position of packages on a conveyor belt. The first neural network uses a conventional target detection neural network to detect unobstructed goods, while the second neural network is optimized for obstructed situations. The results of the first and second neural network detections are allocated using the Hungarian algorithm. The allocated results and the unmatched portions are analyzed to determine the overlap of the packages. The method of the present invention can automatically identify overlapping express parcels in logistics express delivery, facilitating subsequent automatic separation. The method has high recognition efficiency and requires no human intervention, greatly improving the efficiency of logistics transmission.

Description

Technical Field

[0001] The present invention belongs to the technical field of logistics sorting, and in particular relates to a logistics overlap identification method based on deepstream. Background Art

[0002] Although there are single-piece separation systems in the logistics sorting industry, they basically separate single pieces by using the speed difference of the conveyor belt and do not apply actual overlapping recognition technology. Therefore, the following technical problems generally exist:

[0003] 1. Relying solely on the differential speed of the conveyor belt to separate individual pieces has a relatively low accuracy rate.

[0004] 2. The traditional separation method requires manual identification and manual separation of unsuccessful logistics express parcels, which greatly wastes manpower. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a logistics overlap identification method based on deepstream, which can automatically identify overlapping express parcels in logistics express, facilitate subsequent automatic separation, have high recognition efficiency, do not require human participation, and greatly improve the efficiency of logistics transmission.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention provides a logistics overlap recognition method based on DeepStream, including target recognition neural network overlap recognition, wherein the target recognition neural network overlap recognition includes simultaneously using two target detection neural networks to detect the position of packages on a conveyor belt; the first neural network uses a conventional target detection neural network to detect unobstructed goods, and the second neural network is optimized for occlusion situations, by introducing the concept of proposal, applying the NMS algorithm, and citing a loss function for calculation, to simultaneously detect obscured and unobstructed goods; the results of the first and second neural network detections are allocated using the Hungarian algorithm; and the allocated results and the unsuccessfully matched parts are analyzed to obtain the overlap of the packages.

[0008] Furthermore, the concept of proposal includes predicting multiple instances that may be highly overlapping instead of predicting a single instance for each proposal. The loss function uses EMD loss, which is as follows:

[0009]

[0010] Among them, π represents a certain combination of (1,2,...,K), and the kth item is π k , g πk is the πth k detection boxes, L cls and L reg are classification loss and regression loss, c i is the class label, l i are the associated coordinates,

[0011] Minimize the gap between the predicted and proposed box instance sets, improve the NMS algorithm, and perform an additional test before one bounding box suppresses another to determine whether the two detection boxes come from the same proposal. If so, no suppression is performed.

[0012] Furthermore, the recognition method also includes instance segmentation neural network overlapping recognition, which applies a two-layer convolutional network to model the instance segmentation neural network as consisting of two overlapping layers. The top layer detects occluded objects, and the bottom layer detects occluded objects. The two-layer relationship is used to perform clear modeling, naturally decouple the boundaries of occluding and occluded instances, and comprehensively consider the interaction between them in the mask regression process. After the instance segmentation is completed, the occlusion of the object is judged based on the minimum bounding box of the instance.

[0013] The beneficial effects of the present invention are:

[0014] This paper presents a DeepStream-based logistics overlap identification method. By introducing a novel convolutional neural network, the method accepts data before and after additional testing as input, performs a second round of predictions, and corrects possible mispredictions. The instance segmentation network backbone consists of nonlinear convolutional layers and global pooling layers, without feature concatenation or FC layers. This effectively improves the speed of instance segmentation operations and meets the real-time detection requirements.

[0015] Other advantages, objectives and features of the present invention will be described in the following description and will be apparent to those skilled in the art to some extent, or those skilled in the art can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. DETAILED DESCRIPTION

[0016] The present invention provides a logistics overlap recognition method based on DeepStream, including target recognition neural network overlap recognition, wherein the target recognition neural network overlap recognition includes simultaneously using two target detection neural networks to detect the position of packages on a conveyor belt; the first neural network uses a conventional target detection neural network to detect unobstructed goods, and the second neural network is optimized for occlusion situations, by introducing the concept of proposal, applying the NMS algorithm, and citing a loss function for calculation, to simultaneously detect obscured and unobstructed goods; the results of the first and second neural network detections are allocated using the Hungarian algorithm; and the allocated results and the unsuccessfully matched parts are analyzed to obtain the overlap of the packages.

[0017] In this embodiment, the concept of input proposal includes predicting multiple instances that may be highly overlapping instead of a single instance for each proposal suggestion box. This method makes it possible for nearby suggestion boxes to infer the same set of instances. The loss function uses EMD loss, which is as follows:

[0018]

[0019] Among them, π represents a certain combination of (1,2,...,K), and the kth item is π k , g πk is the πth k detection boxes, L cls and L reg are classification loss and regression loss, c i is the class label, l i are the associated coordinates,

[0020] To minimize the gap between the predicted and proposed box instance sets, the NMS algorithm is improved. Before one bounding box suppresses another, an additional test is performed to determine whether the two detection boxes come from the same proposal. If so, no suppression is performed. Although the above algorithm may identify stacked objects, it may lead to an increase in false positives. Therefore, a refinement module is introduced, which is a new convolutional neural network that accepts data before and after the additional test as input to perform a second round of predictions to correct possible incorrect predictions.

[0021] In this embodiment, the recognition method also includes instance segmentation neural network overlapping recognition, which applies a two-layer convolutional network to model the instance segmentation neural network as consisting of two overlapping layers. The top layer detects occluding objects, and the bottom layer detects occluded objects. The two-layer relationship is used to perform clear modeling, naturally decouple the boundaries of occluding and occluded instances, and comprehensively consider the interaction between them in the mask regression process. After the instance segmentation is completed, the occlusion of the object is determined based on the minimum bounding box of the instance. The backbone network of the instance segmentation network is composed of nonlinear convolutional layers and global pooling layers, without using feature cascades or fc layers, which can effectively improve the speed of instance segmentation operations and meet the real-time requirements of detection.

[0022] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A logistics overlap identification method based on deepstream, characterized by: It includes target recognition neural network overlapping recognition, and the target recognition neural network overlapping recognition includes using two target detection neural networks to detect the position of the package on the conveyor belt at the same time; the first neural network uses a conventional target detection neural network to detect unobstructed goods, and the second neural network is optimized for occlusion situations. By introducing the concept of proposal, applying the NMS algorithm, and referencing the loss function for calculation, the detection of occluded and unobstructed goods is performed simultaneously; the results of the first and second neural network detections are assigned using the Hungarian algorithm; the assigned results and the unsuccessfully paired parts are analyzed to obtain the overlapping situation of the package; the introduction of the concept of proposal includes not predicting a single instance for each proposal suggestion box, but predicting multiple instances that may be highly overlapping. The loss function uses EMD loss, and the loss function is shown below Among them, π represents a certain combination of (1,2,...,K), and the kth item is π k , is the πth k detection boxes, L cls and L reg are classification loss and regression loss, c i is the class label, l i are the associated coordinates, Minimize the gap between the predicted instance sets associated with the proposal boxes, improve the NMS algorithm, and perform an additional test before one bounding box suppresses another to determine whether the two detection boxes come from the same proposal. If so, no suppression is performed. The recognition method also includes instance segmentation neural network overlapping recognition. The instance segmentation neural network overlapping recognition applies a two-layer convolutional network to model the instance segmentation neural network as consisting of two overlapping layers. The top layer detects occluded objects and the bottom layer detects occluded objects. The two-layer relationship is used to perform clear modeling, naturally decouple the boundaries of occluded and occluded instances, and comprehensively consider the interaction between them in the mask regression process. After the instance segmentation is completed, the occlusion of the object is judged according to the minimum bounding box of the instance.

Images