A target detection method based on event data

Abstract

The application discloses a target detection method based on event data, which comprises the following steps: converting event stream data of an event camera into event frame data as training sample data; introducing the converted event frame data into a graph neural network structure for training to obtain a trained graph neural network model; converting new event stream data into event frame data, inputting the event frame data into the trained graph neural network model, and obtaining a target detection result of the event frame data; and the target detection result comprising a type of the target and a position of the target in the event frame. The application can identify the type of the target and the position of the target in a scene with a complex texture target and a cluttered background.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, in particular to a target detection method based on event data. BACKGROUND

[0002] Target detection, as one of the popular research directions in the field of computer vision, has a wide range of applications in intelligent monitoring systems, virtual reality, unmanned driving, human-computer interaction and industrial robots. The basic problem to be solved by target detection is to select the target of interest from a given sequence of image frames, and to solve the position of the target in each frame by computer vision algorithm. Generally speaking, the current target detection algorithm is based on the motion sequence of a series of image frames captured by a traditional CMOS camera, and these algorithms have been quite mature.

[0003] With the development of image and computer vision field, its application in all aspects of people's life, people have higher requirements on technology, but the traditional camera has encountered a bottleneck, which is a kind of visual sensor with fixed frame output image, which inevitably faces the problems of delay, power consumption and other aspects, especially in high dynamic, high resolution environment, there are obvious limitations, which affect the application in object recognition, segmentation, tracking and optical flow estimation and other task algorithm research. Therefore, a new type of visual sensor appears, which is inspired by the biological retina, based on event sensor, which provides continuous output asynchronous event stream. The event triggered by the asynchronous work of the pixel only outputs the address and information of the pixel whose light intensity changes, instead of passively reading the information of each pixel in the frame. Therefore, with the characteristics of scene change, image super sparse representation and real-time high dynamic response of event asynchronous output, it can be widely used in high-speed target tracking, real-time monitoring and robot vision.

[0004] Early detection algorithms focus on the low delay and low processing characteristics of event-driven vision systems, and detect simple moving targets by clustering blocks, circles or lines from event sources. Subsequently, many researchers use iterative nearest point algorithm, Monte-Carlo, particle filter algorithm and other event-by-event adaptive detection methods to detect complex, high-contrast predefined shapes. The above-mentioned event stream-based detection algorithms are all for targets without complex texture and scenes without cluttered background, but there are few algorithms for target bounding box detection in complex scenes. Recently, event data target detection technology based on deep learning has become a research hotspot, and the main research direction is to process event information in SNN, CNN, GCN and other networks, which can further dig the characteristics of event camera on the basis of handcrafted framework and improve the effect, but most of these algorithms convert event stream data into event frame form by adaptive sampling method, and use traditional convolutional neural network to extract event frame features and obtain target detection frame.

[0005] There are currently two processing methods for processing event stream data of an event camera. One is a fixed sampling time interval method, and the other is a fixed event number sampling method. The fixed time interval method collects, processes and analyzes event stream data of a time period as a whole, and the number of events processed in each time interval may be different, which may cause the generated image texture to be too sparse or too dense, resulting in unclear images. The fixed event number sampling method produces event frames as the fixed time interval method does. In the case that objects in a scene have different motion speeds, the fixed event number method may cause the generated image texture to be too sparse, or the generated image texture to be too dense, resulting in motion blur in the image. SUMMARY

[0006] In view of this, the present application provides a target detection method based on event data, which converts event stream data into an adaptive time-number sampling method in the form of event frames, and designs a graph neural network (GNN) target detection algorithm based on event frames, and can identify the type and position of the target in a complex texture target and cluttered background scene.

[0007] The present application discloses a target detection method based on event data, which comprises:

[0008] Step 1: converting event stream data of an event camera into event frame data as training sample data;

[0009] Step 2: importing the converted event frame data into a graph neural network structure for training to obtain a trained graph neural network model;

[0010] Step 3: converting new event stream data into event frame data according to step 1, inputting the event frame data into the trained graph neural network model, and obtaining a target detection result of the event frame data; the target detection result includes the type of the target and the position of the target in the event frame.

[0011] Further, the step 1 comprises:

[0012] Step 11: event stream data representation modeling;

[0013] Step 12: generating event frames based on data sampling.

[0014] Further, the step 11 comprises:

[0015] The data output by the event camera is represented as event stream data, which is a sequence composed of a series of event points. Each event point contains three elements: a timestamp , a pixel coordinate and a polarity; the timestamp time at which the event occurred, pixel coordinates position of the pixel at which the event occurred in the image coordinate system; polarity indicates the type of the event point

[0016] The event stream data is represented as a sequence:

[0017] (1)

[0018] wherein, represents the pixel position of the i-th event, represents the horizontal coordinate in the pixel position, represents the vertical coordinate in the pixel position, represents the time at which the i-th event occurred, represents the polarity of the i-th event.

[0019] Further, the timestamps of the event stream data are monotonically increasing, i.e.

[0020] If a fixed time range is given, the event stream data generated by the event camera in the fixed time range is represented as:

[0021] (2)

[0022] wherein, represents the number of event data in the fixed time range, represents the i-th event stream data, i.e. represents the set of event stream data in the fixed time range.

[0023] Further, the event stream data is represented as a point set in a four-dimensional manifold spanned by coordinate axis , coordinate axis , time axis and polarity axis The point set of positive polarity event points , the point set of negative polarity event points , and the event stream data of arbitrary polarity are compressed into three-dimensional space by event field:

[0024] (3)

[0025] wherein, is the Kronecker function, is the impulse function delay ​​​​​​​​a pulse function after a certain time interval;

[0026] and are respectively represented as:

[0027] (4)

[0028] (5).

[0029] Further, the step 12 comprises:

[0030] Step 121: data preprocessing: extracting all timestamp information of the event camera event stream data and normalizing it to adapt to the input requirements of the LSTM model;

[0031] Step 122: LSTM model training: constructing an LSTM model, training the LSTM model using timestamp data, and optimizing the LSTM model performance by adjusting the hyperparameters of the model; the LSTM model includes an input layer, an LSTM layer, and an output layer; the hyperparameters include learning rate, hidden layer size;

[0032] Step 123: time series prediction: using historical timestamp data to predict future timestamp data ; wherein, represents the timestamp data, i.e., the last timestamp of the current sampling; is the timestamp data, is the timestamp data, is the timestamp data;

[0033] Step 124: calculating , and the time scaling factor , according to the time scaling factor ， to obtain the sampling time interval;

[0034] Step 125: data sampling: sampling according to the new sampling time interval, assuming that the last timestamp of the current sampling is , then let , update the current sampling timestamp, and update the historical timestamp data used for prediction; then repeat step 123, and cycle in turn until the sampling is completed.

[0035] Further, the step 124 comprises:

[0036] According to the historical timestamp data and the future timestamp data, a mean value of the historical timestamp is obtained and a mean value of the future timestamp ;

[0037] According to the mean value of the historical timestamp and the mean value of the future timestamp , a time scaling factor is obtained ;

[0038] According to the time scaling factor and a current sampled time interval , a next sampled time interval is obtained .

[0039] Further, expressions of the mean value of the historical timestamp , the mean value of the future timestamp , and the time scaling factor are respectively:

[0040] (6)

[0041] wherein, the mean value of the historical timestamp is denoted as the mean value of the future timestamp is denoted as

[0042] If the time interval of the i-th sampling is , then the time interval of the i+1-th sampling is .

[0043] Further, the graph neural network structure comprises a backbone network, a classifier and a regressor; output ends of the backbone network are connected with the classifier and the regressor respectively; the backbone network comprises a first branch, a second branch and a third branch; the first branch, the second branch and the third branch are connected in sequence, and output ends of the third branch are connected with the classifier and the regressor respectively;

[0044] The first branch comprises graph convolution GCNConv, rectified linear unit activation function Relu, batch normalization layer BatchNorm, dropout layer Dropout, graph convolution GCNConv, rectified linear unit activation function Relu, batch normalization layer BatchNorm and maximum pooling layer Max Pool connected in sequence;

[0045] The second branch comprises sequentially connected spline convolution SplineConv, rectified linear unit activation function Relu, batch normalization layer BatchNorm, dropout layer Dropout, spline convolution SplineConv, rectified linear unit activation function Relu, batch normalization layer BatchNorm, and maximum pooling layer Max Pool;

[0046] The third branch comprises sequentially connected graph convolution GCNConv, rectified linear unit activation function Relu, batch normalization layer BatchNorm, dropout layer Dropout, graph convolution GCNConv, rectified linear unit activation function Relu, batch normalization layer BatchNorm, and global mean pooling layer Global Mean Pool;

[0047] The classifier comprises sequentially connected linear layer Linear, rectified linear unit activation function Relu, linear layer Linear, and normalized exponential function Softmax.

[0048] The regressor comprises sequentially connected linear layer Linear, rectified linear unit activation function Relu, linear layer Linear, and S-shaped growth curve activation function Sigmoid.

[0049] The output end of the third branch is connected with the linear layer Linear in the classifier and the regressor, respectively.

[0050] Further, in the backbone network, the input of the first GCNConv of the first branch is event stream data S n node information and edge index, wherein the node information is input in the form of an event frame, the input of the remaining GCNConv is the output of the previous network layer and the edge index information; the input of all SplineConv is the output of the previous network layer, the edge index and the attribute information of the edge; the input of all Dropout and BatchNorm is the output of the previous network layer, wherein the dropout probability of Dropout is a positive number less than 1; the input of all Max Pool and Global Mean Pool is the output of the previous network layer plus batch node information. represents the horizontal coordinate in the pixel position, represents the vertical coordinate in the pixel position, represents the time when the i-th event occurs, represents the polarity of the i-th event.

[0051] ​​In the detection head head, the Softmax and Sigmoid are used to output the predicted target class probability and target bounding box data respectively, and the output result of target detection is formed.

[0052] The loss function adopts a weighted sum of classification loss and bounding box regression loss, wherein the classification loss adopts cross-entropy loss Cross-Entropy loss, and the regression loss adopts smooth L1 loss Smooth-L1 loss, and therefore the loss function Loss is represented as: , 、 is a weight coefficient, and .

[0053] Due to the adoption of the above technical solutions, the present application has the following advantages: the present application utilizes the powerful time series data prediction capability of the LSTM neural network, and through the prediction of the timestamp data, adaptive time-data sampling is completed, which can effectively solve the problem that the existing equal-interval time sampling may cause the generated image texture to be too sparse or too dense, resulting in unclear images. At the same time, the present application also proposes a new GNN neural network architecture to solve the target detection problem of the target of interest in event data under the premise of ensuring real-time performance. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments described in the embodiments of the present application, and other drawings can also be obtained by those skilled in the art according to these drawings.

[0055] Figure 1 FIG. 1 is a schematic diagram of an event image target detection method architecture based on a graph neural network according to an embodiment of the present application;

[0056] Figure 2 FIG. 2 is a data sampling flowchart for converting event stream data of a camera into an event frame according to an embodiment of the present application;

[0057] Figure 3 FIG. 3 is a graph neural network structure diagram according to an embodiment of the present application. DETAILED DESCRIPTION

[0058] The present application is further described in conjunction with the drawings and embodiments, and the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those skilled in the art should belong to the scope of protection of the embodiments of the present application.

[0059] Referring to Figure 1, the present application provides an embodiment of a target detection method based on event data, which includes:

[0060] Step 1: Convert the event stream data of the event camera into event frame data as training sample data;

[0061] Step 2: Import the converted event frame data into the graph neural network structure for training to obtain a trained graph neural network model;

[0062] Step 3: Convert the new event stream data into event frame data according to step 1, input the event frame data into the trained graph neural network model, and obtain the target detection results of the event frame data; the target detection results include the type of target and the position of the target in the event frame.

[0063] Optionally, step 1 includes:

[0064] Step 11: Modeling event stream data representation;

[0065] Step 12: Generate event frames based on data sampling.

[0066] Optionally, step 11 includes:

[0067] The data output by the event camera is represented as event stream data. The event stream data consists of a series of event points. Each event point contains three elements: timestamp , pixel coordinates and polarity; timestamp Indicates the time when the event occurred, pixel coordinates Indicates the position of the pixel that generated the event in the image coordinate system; polarity indicates the type of event point, which usually has two values: positive and negative, indicating that the light intensity of the pixel that should not generate the event is stronger or weaker than before;

[0068] Event stream data is represented as a sequence:

[0069] (1)

[0070] in, Indicates the The pixel location of the event, represents the horizontal coordinate of the pixel position, represents the vertical coordinate in the pixel position, Indicates the The time when the event occurred, Indicates the The polarity of an event.

[0071] Optionally, the timestamp of the event stream data is monotonically increasing, that is ;

[0072] If a fixed time range is given, the event stream data generated by the event camera in the fixed time range is represented as:

[0073] (2)

[0074] wherein, represents the number of event data in the fixed time range, is the i-th event stream data, i.e. , , represents the set of event stream data in the fixed time range.

[0075] Optionally, the event stream data is represented as a point set in a four-dimensional manifold spanned by coordinate axis , , and polarity axis , , , The event field is compressed into a three-dimensional space :

[0076] (3)

[0077] wherein, is a Kronecker function, is an impulse function delayed by ;

[0078] and are respectively represented as:

[0079] (4)

[0080] (5).

[0081] Optionally, the step 12 comprises:

[0082] The present application implements an adaptive sampling method, which utilizes the powerful time series data prediction perception of the LSTM network, intelligently adjusts the sampling interval, ensures that more event stream data of a frame of data is sampled when events occur frequently, and samples more sparsely when events are sparse, with a longer time interval, reduces unnecessary sampling times, and thus reduces storage and computing costs. The sampling steps for generating a specific event frame are as follows:

[0083] Step 121: data preprocessing: extract all timestamp information of event camera event stream data, and normalize it to adapt to the input requirements of the LSTM model;

[0084] Step 122: LSTM model training: build an LSTM model, train the LSTM model using timestamp data, and optimize the performance of the LSTM model by adjusting the hyperparameters of the model; the LSTM model includes an input layer, an LSTM layer, and an output layer; the hyperparameters include learning rate, hidden layer size;

[0085] Step 123: time series prediction: using historical timestamp data , predict future timestamp data ; wherein, represents the timestamp data, i.e. the last timestamp of the current sampling; is the timestamp data, is the timestamp data, is the timestamp data;

[0086] Step 124: calculate , and time scaling factor , according to the time scaling factor ， get the sampling time interval;

[0087] Step 125: data sampling: sampling according to the new sampling time interval, assuming the last timestamp of the current sampling is , let , update the current sampling timestamp, and update the historical timestamp data used for prediction; then repeat step 123, and cycle in turn until the sampling is completed, as shown in Figure 2 .

[0088] Optionally, the step 124 includes:

[0089] According to the historical timestamp data and the future timestamp data, get the mean value of the historical timestamp and the mean value of the future timestamp;

[0090] According to the mean value of the historical timestamp and the mean value α of the future timestamp, get the time scaling factor ;

[0091] According to the time scaling factor and the time interval of the current sampling , the time interval of the next sampling is obtained .

[0092] Optionally, the expression of the mean of the historical timestamps , the mean of the future timestamps , the time scaling factor is respectively:

[0093] (6)

[0094] wherein, represents the mean of the historical timestamps, represents the mean of the future timestamps;

[0095] If the time interval of the i-th sampling is , the time interval of the i+1-th sampling is .

[0096] Optionally, referring to Figure 3 , the network architecture utilizes GCNConv (graph convolution), SplineConv (spline convolution), and pooling layer network structures to form a backbone network, and respectively connects a classifier and a regressor after the backbone network, so as to realize the target detection function of event data.

[0097] The graph neural network structure comprises a backbone network, a classifier and a regressor; output ends of the backbone network are connected with the classifier and the regressor respectively; the backbone network comprises a first branch, a second branch and a third branch; the first branch, the second branch and the third branch are connected in sequence, and output ends of the third branch are connected with the classifier and the regressor respectively;

[0098] The first branch comprises graph convolutions GCNConv, a rectified linear unit activation function Relu, a batch normalization layer BatchNorm, a dropout layer Dropout, graph convolutions GCNConv, a rectified linear unit activation function Relu, a batch normalization layer BatchNorm and a maximum pooling layer Max Pool connected in sequence;

[0099] The second branch comprises spline convolutions SplineConv, a rectified linear unit activation function Relu, a batch normalization layer BatchNorm, a dropout layer Dropout, spline convolutions SplineConv, a rectified linear unit activation function Relu, a batch normalization layer BatchNorm and a maximum pooling layer Max Pool connected in sequence;

[0100] The third branch includes the graph convolution GCNConv, the rectified linear unit activation function Relu, the batch normalization layer BatchNorm, the dropout layer Dropout, the graph convolution GCNConv, the rectified linear unit activation function Relu, the batch normalization layer BatchNorm, and the global mean pooling layer Global Mean Pool connected in sequence;

[0101] The classifier includes a linear layer Linear, a rectified linear unit activation function Relu, a linear layer Linear, and a normalized exponential function Softmax connected in sequence;

[0102] The regressor includes a linear layer Linear, a rectified linear unit activation function Relu, a linear layer Linear, and an S-shaped growth curve activation function Sigmoid connected in sequence;

[0103] The output end of the third branch is connected to the linear layer in the classifier and the regressor respectively.

[0104] Optionally, in the backbone network, the input of the first GCNConv of the first branch is the event stream data S n Node information and edge indexes, where node information is input in the form of event frames, and the inputs of the remaining GCNConvs are the output of the previous network layer and edge index information; the inputs of all SplineConvs are the output of the previous network layer, edge indexes, and edge attribute information; the inputs of all Dropout and BatchNorms are the output of the previous network layer, where the dropout probability of Dropout is a positive number less than 1; the inputs of all Max Pools and Global Mean Pools are the output of the previous network layer plus batch node information; represents the horizontal coordinate of the pixel position, represents the vertical coordinate in the pixel position, Indicates the The time when an event occurs, Indicates the The polarity of an event;

[0105] In the detection head, Softmax and Sigmoid are used to output the predicted target category probability and target bounding box data respectively to form the output result of target detection;

[0106] The loss function adopts the weighted sum of classification loss and bounding box regression loss, where the classification loss adopts cross-entropy loss and the regression loss adopts smooth L1 loss. Therefore, the loss function Loss is expressed as: , 、 are weight coefficients, and .

[0107] Finally, it should be noted that the above examples are merely intended to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the above examples, those of ordinary skill in the art should understand that the specific implementation of the present application can still be modified or equivalently replaced without departing from the spirit and scope of the present application, and any modification or equivalent replacement that does not depart from the spirit and scope of the present application should be covered within the protection scope of the claims of the present application.

Claims

1. A method for target detection based on event data, characterized in that, The method comprises the following steps: Step 1: converting event stream data of an event camera into event frame data as training sample data; Step 2: importing the converted event frame data into a graph neural network structure for training to obtain a trained graph neural network model; Step 3: converting new event stream data into event frame data according to step 1, and inputting the event frame data into the trained graph neural network model to obtain a target detection result of the event frame data; The target detection result comprises a type of the target and a position of the target in the event frame; The step 1 comprises: Step 11: event stream data representation modeling; Step 12: generating an event frame based on data sampling; The step 12 comprises: Step 121: data preprocessing: extracting all timestamp information of the event stream data of the event camera, and performing normalization processing on the timestamp information to adapt to input requirements of an LSTM model; Step 122: LSTM model training: constructing an LSTM model, training the LSTM model using the timestamp data, and optimizing the performance of the LSTM model by adjusting hyperparameters of the model; the LSTM model comprises an input layer, an LSTM layer and an output layer; the hyperparameters comprise a learning rate and a hidden layer size; Step 123: Time series prediction: using historical timestamp data , predict future timestamp data ; wherein, represents the timestamp data, i.e. the last timestamp of the current sampling; is the timestamp data, is the timestamp data, is the timestamp data; Step 124: Calculate the mean of the history timestamps , the mean of the future timestamps and the time scaling factor , according to the time scaling factor ， get the sampling time interval; if the time interval of the ith sampling is , then the sampling time interval of the i+1th is ; Step 125: Data sampling: sampling according to the new sampling time interval, assuming that the last timestamp of the current sampling is Let , update the current sampling timestamp, and update the prediction historical timestamp data ; then repeat step 123, in turn, until the sampling is completed; The graph neural network structure comprises a backbone network, a classifier and a regressor; output ends of the backbone network are connected with the classifier and the regressor respectively; the backbone network comprises a first branch, a second branch and a third branch; the first branch, the second branch and the third branch are connected in sequence, and output ends of the third branch are connected with the classifier and the regressor respectively; The first branch comprises graph convolution GCNConv, a rectified linear unit activation function Relu, a batch normalization layer BatchNorm, a dropout layer Dropout, graph convolution GCNConv, a rectified linear unit activation function Relu, a batch normalization layer BatchNorm and a maximum pooling layer Max Pool which are connected in sequence; The second branch comprises spline convolution SplineConv, a rectified linear unit activation function Relu, a batch normalization layer BatchNorm, a dropout layer Dropout, spline convolution SplineConv, a rectified linear unit activation function Relu, a batch normalization layer BatchNorm and a maximum pooling layer Max Pool which are connected in sequence; The third branch comprises graph convolution GCNConv, a rectified linear unit activation function Relu, a batch normalization layer BatchNorm, a dropout layer Dropout, graph convolution GCNConv, a rectified linear unit activation function Relu, a batch normalization layer BatchNorm and a global mean pooling layer Global Mean Pool which are connected in sequence; The classifier comprises a linear layer Linear, a rectified linear unit activation function Relu, a linear layer Linear and a normalized exponential function Softmax which are connected in sequence; The regressor comprises a linear layer Linear, a rectified linear unit activation function Relu, a linear layer Linear and a sigmoid growth curve activation function Sigmoid which are connected in sequence; The output ends of the third branches are connected with linear layers Linear in the classifier and the regressor respectively.

2. The method of claim 1, wherein, The step 11 comprises: The data output by the event camera is represented as event stream data, which consists of a sequence of event points, each of which contains three elements: a timestamp , a pixel coordinate , and a polarity; the timestamp represents the time at which the event was generated, the pixel coordinate represents the location of the pixel at which the event was generated in the image coordinate system; the polarity represents the type of event point; The event stream data is represented as a sequence: (1) wherein, represents a pixel position of the th event, represents a horizontal coordinate in the pixel position, represents a vertical coordinate in the pixel position, represents a time of the th event, represents a polarity of the th event.

3. The method of claim 2, wherein, Timestamps of event stream data are monotonically increasing, i.e. ; If a fixed time range is given, the event stream data generated by the event camera in the fixed time range is represented as: (2) wherein, represents the number of event data in the fixed time range, is the event stream data, i.e. , represents the set of event stream data in the fixed time range.

4. The method of claim 3, wherein, Event stream data is represented as a set of points in a four-dimensional manifold spanned by coordinate axes , coordinate axis , time axis , and polarity axis . Positive polarity event point set , negative polarity event point set , arbitrary polarity event stream data is compressed into three-dimensional space by event field : (3) wherein is the Kronecker function, is the impulse function delay the impulse function after and are represented by: (4) (5)。 5. The method of claim 1, wherein, The step 124 comprises: According to the historical timestamp data and the future timestamp data, a mean value of the historical timestamp is obtained and a mean value of the future timestamp ; the mean of the historical timestamps and the mean of the future timestamps resulting in a time scaling factor ; According to the time scaling factor and the current sampled time interval to obtain the next sampled time interval .

6. The method of claim 5, wherein, the mean of the historical timestamps the mean of the future timestamps a time scaling factor are respectively (6) wherein, represents the mean of the historical timestamps, represents the mean of the future timestamps.

7. The method of claim 1, wherein, In the backbone network, the input of the first GCNConv of the first branch is the event stream data S n Node information and edge index, where node information is input in the form of event frames, and the input of other GCNConv is the output of the previous network layer and edge index information; the input of all SplineConv is the output of the previous network layer, edge index and edge attribute information; the input of all Dropout and BatchNorm is the output of the previous network layer, where the drop probability of Dropout is a positive number less than 1; the input of all Max Pool and GlobalMean Pool is the output of the previous network layer plus batch node information; represents the horizontal coordinate of the pixel position, represents the vertical coordinate in the pixel position, Indicates the The time when an event occurs, Indicates the The polarity of an event; In the detection head head, the predicted target class probability and the target bounding box data are output by using Softmax and Sigmoid respectively, and the output result of target detection is formed. The loss function adopts a weighted sum of a classification loss and a bounding box regression loss, wherein the classification loss adopts a Cross-Entropy loss, and the regression loss adopts a Smooth-L1 loss, and thus the loss function Loss is represented as: , 、 is a weight coefficient, and .

Images