Lightweight pointpillars autonomous driving 3D object detection method based on column self-attention

Abstract

The application discloses a kind of lightweight PointPillars automatic driving 3D target detection methods based on column self-attention, obtains point cloud dataset, and the point cloud image of surrounding 360 is cropped to image visual angle;Using the encoder network structure of transformer to the feature extraction of point cloud data, then through multilayer perceptron, to the dimensionality of point cloud feature is upgraded;Point cloud data is cut into column, and these columns are distributed to the original world, form 2D pseudo image;Afterwards, pseudo image is sent into 2D main network and carries out 2D convolution operation, then 2D feature map is obtained and is sent into the detection head based on SSD, finally input predicted result, and the prediction result includes the 3D center point coordinate and length-width-height of object and steering angle;Transformer technology and depth separable convolution technology are introduced in the feature extraction part in the application, improve the extraction ability of network to context semantic feature, and greatly reduce the parameter quantity of network, realize lightweight.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of autonomous driving cars, and particularly relates to a lightweight PointPillars autonomous driving 3D target detection method based on columnar self-attention. BACKGROUND

[0002] In recent years, the development trend of technology is to introduce automation and intelligence into life, thereby improving the quality of people's life. With the rapid development of social economy and science and technology, cars have become popular and gradually become one of the necessary means of travel for people. Whether it is commuting or traveling on holidays, cars are closely related to our life. Accordingly, while enjoying the great convenience brought to us by cars, we inevitably have to face some problems. Long-time high-intensity commuting pressure is easy to cause driving fatigue, and the driver's attention is difficult to concentrate, and the driving experience is also poor. And it greatly increases the probability of traffic accidents. Among them, the accidents caused by human factors account for a large proportion of the accidents, such as driving skills, mental changes and fatigue, which increase the risk of car accidents. Autonomous driving technology is considered one of the most promising remedies to protect humans from serious collisions and receives more and more attention. The key to autonomous driving is perception, especially 3D target detection, which is the core foundation of the perception system and is crucial for path planning, motion prediction, collision avoidance, etc.

[0003] The core of autonomous driving cars is to be intelligent like humans, and the most important thing is environmental perception. 3D object detection plays a crucial role in Autonomous Vehicles (AVs). Laser radar and color camera are the main sensors equipped on AVs. In target detection, it is mainly divided into two categories: 2D target detection and 3D target detection. 2D target detection can only obtain the position information of the target in the two-dimensional image, and the 2D size obtained by only 2D target detection cannot complete this target. In contrast, 3D target detection can obtain the position, size, heading angle, etc. of the object in three-dimensional space, so as to calculate the center point of the bottom of the vehicle. 3D target detection mainly relies on devices such as laser radar, binocular camera and monocular camera.

[0004] At present, most powerful 3D detectors rely on laser radar because the device can provide accurate depth information according to three-dimensional point cloud. However, there are some limitations in the laser radar-based system device. In bad weather (such as rain, snow, fog), the perception ability of laser radar to the environment will be reduced, and incomplete or inaccurate data will be generated. Secondly, when detecting objects with different surface reflectivity (such as mirrors, glass, water, etc.), laser radar may miss or misdetect. SUMMARY

[0005] In view of the deficiencies of the prior art described above, the purpose of the present application is to provide a lightweight PointPillars autonomous driving 3D target detection method based on column self-attention, aiming to solve the limitations and defects existing in the prior art in laser radar detection, resulting in inaccurate 3D target data detection, large parameter quantity, and the problem that the 3D target detection method cannot be lightweight.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0007] A lightweight PointPillars autonomous driving 3D target detection method based on column self-attention, comprising:

[0008] Step one: obtain point cloud data, image data and device parameters of public roads, wherein the device parameters include camera internal and external parameters, laser radar origin coordinates and labeled ground true bounding box;

[0009] Step two: crop the point cloud data, and crop the 360° point cloud around the vehicle body to only the point cloud appearing in the camera view;

[0010] Step three: then input the processed original point cloud into the first improved PillarsFeature Net module of PointPillars, and increase the local transformer structure in the PillarsFeature Net module to improve the feature extraction capability, and finally obtain a 2D pseudo image;

[0011] Step four: send the pseudo image obtained in the above step into the deep separable lightweight backbone network based on channel attention for feature extraction to obtain a 2D feature map;

[0012] Step five: send the 2D feature map into the target detection head based on SSD to obtain the detection result;

[0013] Step six: mark the detection result to the original point cloud to realize the 3D target detection task of the autonomous driving vehicle in the 3D world.

[0014] Further, the local self-attention transformer process of the PillarsFeature Net module in step three comprises:

[0015] Step 3-1: Obtain the cropped raw point cloud data obtained in step two, which contains (X, Y, Z, R) of all points in the scene appearing in the camera field of view, where X, Y, Z and R represent the three-dimensional world coordinates of a single point and the surface reflection intensity, respectively, which are established in the laser radar coordinate system, with the front of the car as the x-axis, the left side of the car as the y-axis, and the top of the car as the z-axis;

[0016] Step 3-2: Point cloud columnization, the obtained raw point cloud data is sent to the voxelize module for cutting, cutting in the xoy plane, and specifying the size of the cutting column as [0.16, 0.16, 4], and calculating the corresponding coordinates of each column in the original three-dimensional world as the return value to the PillarsFeature Net module;

[0017] Step 3-3: Feature dimensionality, after obtaining the segmented column, the cluster center of all points in each column, the center point offset of each column, and the distance from the observer to the point are calculated, all of which are taken as the features of the point, and finally a point with a feature of (4+6) = 10 is obtained;

[0018] Step 3-4: Then it is sent to the local self-attention transformer for local feature extraction and attention calculation; after the above step, a tensor with a shape such as [batchsize*6542, 32, 10] is obtained, where the first dimension represents the number of columns in a sample in this batch, the second dimension represents the number of points contained in each column, and the third dimension represents the specific feature information of each point; the tensor is sent to the encoder structure of the transformer, which is first positionally encoded, and the encoding formula is:

[0019]

[0020] Where PE is the result of position encoding, p is the current point, i represents the current dimension, d f represents the total feature dimension;

[0021] Then calculate the self-correlation of each column:

[0022]

[0023] Where Q, K, and V represent the query matrix, key matrix, and value matrix, represents the scaling ratio to prevent the calculation result from exceeding the limit; the final result Atten(Q, K, V) represents the importance of each point relative to other points;

[0024] Multi-head attention mechanism is adopted:

[0025] MultiHead(Q, K, V) = Concat(head1,..., head n )W t ;

[0026]

[0027] where head1,..., headH are different subvectors, are weights of the subvectors, and then all the subvectors are concatenated together via concat, and finally a weight W t is applied to obtain the final attention result; then a residual structure and a forward propagation module are applied, and multiple iterations are performed;

[0028] Step 3-5: Finally, the point cloud features obtained through the encoder are restored to 2D pseudo images according to the relative coordinates.

[0029] Further, the feature extraction network in step four is characterized in that it comprises:

[0030] After obtaining the 2D pseudo image in step 3, a lightweight network is redesigned.

[0031] The size of the input image is [64, 496, 432], and through three times of downsampling, three feature maps with different scales are saved, which are [64, 248, 216], [128, 124, 108], and [256, 62, 54], respectively. Then, different transposed convolutions are applied to restore the unified size [384, 124, 108], and finally, the feature maps containing different scale features are obtained by concatenation.

[0032] Among them, the convolution of the network adopts deep separable convolution to reduce the parameter quantity of the model, and the Bneck structure adopts a channel-based attention mechanism to improve the feature extraction capability of the model; and the Bneck structure contains an SEBlock channel attention module, which mainly functions to enable the network to focus on more important feature layers; at the same time, a residual structure is adopted to prevent gradient vanishing and gradient explosion.

[0033] Further, in step four, through deep separable convolution, all parameters of the convolution layer are reduced from h o ·w o ·a o 2 ·d o ·d o to h o ·w o ·(a o 2 +d o )·d oThe total parameter amount is 24.3% of the original model.

[0034] Further, the SSD detection head in step five detects includes:

[0035] Step 5-1: The 2D feature map obtained in step four is first passed through three different prediction heads, respectively, a classification head, a regression head, and a direction prediction head; different prediction results are obtained for different prediction heads, for the classification prediction head, the classification category is 3 categories, and cross entropy loss is calculated; for the regression prediction head, the position information of the 3D object is mainly predicted, specifically, the coordinate information of 7 objects is predicted; for the direction prediction head, a binary cross entropy loss is used for calculation;

[0036] Step 5-2: For the ground truth label, each object has 7 parameters, respectively, [x, y, z, w, l, h, θ], wherein the first three represent x, y, z in the laser radar coordinate system, w, l, h represent width, length, and height, respectively, and θ represents the heading angle;

[0037] The SmoothL1 regression loss is used between the real ground label and the prediction box, and the formula of the regression loss is:

[0038]

[0039] Δθ=sin(θ t -θ a );

[0040] Wherein x t y t z t represents the ground truth label, x a y a z a represents the prediction box, w t l t h t represents the width, length, and height of the label, w a l a h a represents the predicted width, length, and height; θ t θ a respectively represent the turning angle of the label and the predicted turning angle;

[0041] The overall regression loss is:

[0042]

[0043] Considering that in a three-dimensional environment, the positioning loss cannot identify the angle of two completely opposite positioning frames, a softmax heading angle classification loss is used to learn the heading angle of the three-dimensional object.

[0044] The object classification loss used is focal loss:

[0045] L cls =-α a (1-p a ) γ log p a :

[0046] In the formula, the classification probability score is represented by p a , where the hyperparameters α a = 0.25 and γ = 2;

[0047] The total loss is represented as:

[0048]

[0049] where N pos represents the number of anchor boxes assigned as positive samples, and weights are again added for different losses, because in 3D object detection, more attention is paid to the coordinate information of the object in the 3D world, so β loc = 2, β cls = 1, and β dir = 0.2.

[0050] Further, the prediction results obtained in step 5-2 are evaluated:

[0051] Generally, the main performance parameter of 3D object recognition is average precision (AP); it is slightly modified in the KITTI benchmark, and the interpolation API R40 of the KITTI standard is used to evaluate the performance of the model:

[0052]

[0053] where R = [0, 1 / 40, 2 / 40, …, 1] represents different recall rate thresholds, and P represents the precision at different recall rates; different IoU thresholds (0.7, 0.5, 0.5) correspond to cars, pedestrians, and cyclists, respectively, and low-confidence prediction boxes are filtered out; finally, the non-maximum suppression NMS is used to retain the prediction box with the highest confidence.

[0054] Further, in step six, the detection results obtained in step five are labeled into the original point cloud to generate a new point cloud with detection results, the prediction results are visualized, and the 3D object detection task of an autonomous vehicle in a 3D world is realized.

[0055] The technical solutions adopted by the application have the following beneficial effects:

[0056] In the present application, the transformer technology and the depth separable convolution technology are introduced in the feature extraction part, the Pillars Feature Net module in PointPillars is improved, the local transformer structure is added in the Pillars Feature Net module, the network extraction ability for context semantic features can be improved, and the inaccuracy of 3D target data detected by the laser radar is avoided; through the depth separable convolution technology, the parameter quantity of the network can be greatly reduced, and the lightweight is realized, so as to realize the 3D target detection in the autonomous vehicle. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 It is the original PointPillars model diagram;

[0058] Figure 2 It is the overall flowchart of the present application;

[0059] Figure 3 It is the detailed structure of the MSA (local point attention) part in the overall flowchart of the present application;

[0060] Figure 4 It is the 2D backbone network flowchart;

[0061] Figure 5 It is the Bneck structure and channel attention module schematic diagram;

[0062] Figure 6 It is the depth separable convolution schematic diagram. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical scheme and effect of the present application more clear and definite, the present application will be further described in detail below with reference to the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application.

[0064] Example 1:

[0065] The present application provides a lightweight PointPillars autonomous driving 3D target detection method based on cylinder self-attention, and the original PointPillars is as shown in Figure 1 In the present application: two modules are added in PointPillars, as shown in Figure 2As shown, first, a local transformer is added inside each column, located in the Point Cloud MSA, and then a deep separable convolution and channel attention mechanism are used in the Backbone. The former improves the model's accuracy by adding a local transformer encoder structure inside each column after columnarization of the point cloud, and the latter reduces the model's parameter amount by using a complete deep separable convolution and channel attention mechanism. The steps include:

[0066] Step one: Obtain point cloud data, image data, and some parameters of the device of the public road.

[0067] First, the original point cloud data, camera parameters, laser radar parameters, and labeling information are obtained from the KITTI website. Then the data is divided into a training set and a test set, with 3712 training sets and 3769 test sets.

[0068] Step two: Crop the point cloud, and crop the 360° point cloud around it to only the point cloud appearing in the camera view.

[0069] Step three: Then input the processed original point cloud into the first improved PillarsFeature Net module of PointPillars, innovatively add a local transformer structure to improve the feature extraction capability, and finally get a 2D pseudo image.

[0070] The cropped original point cloud data obtained in step three contains (X, Y, Z, R) of all points appearing in the camera view in a scene, where the three-dimensional world coordinates and surface reflectivity of a single point are represented in order, and the coordinates are in the laser radar coordinate system. The front of the car is the x-axis, the left side is the y-axis, and the upper side is the z-axis. Then the point cloud is columnarized, and the obtained original point cloud data is sent to the voxelize module for cutting. It is cut in the xoy plane, not in the z-axis direction, and the size of the cutting column is specified as [0.16, 0.16, 4], which is the size of a column. The corresponding coordinates of each column in the original three-dimensional world are calculated and returned as the return value.

[0071] In step three, the features are also upgraded. Specifically, after obtaining some columnarized columns, the cluster center of all points, the center point offset of each column, and the observer-to-point distance are calculated as features of the point, and finally a point with a feature of (4+6) = 10 is obtained. Then it will enter the transformer structure for feature extraction, such as Figure 3The detailed structure of the transformer is shown, which specifically adopts position encoding and self-attention mechanism, and the formula is as follows:

[0072]

[0073] Where PE is the result of position encoding, p is the current point, i represents the current dimension, d f represents the total feature dimension.

[0074] Then calculate the autocorrelation of each column, and the formula used is:

[0075]

[0076] In the formula, Q, K, and V represent the query matrix, the key matrix, and the value matrix, respectively, which represents the scaling ratio to prevent the calculation result from exceeding the limit. The final result Atten(Q, K, V) represents the importance of each point relative to other points. A multi-head attention mechanism is also used, and the formula is:

[0077] MultiHead(Q, K, V) = Concat(head1, …, head n )W t

[0078]

[0079] In the formula, head1... represents different sub-vectors, respectively, which are the weights of the sub-vectors. Then all the sub-vectors are concatenated together, and finally a weight W t is applied to obtain the final attention result. Then it passes through a residual structure and a forward propagation module, and this process is iterated multiple times.

[0080] Step three: Finally, map the 2D pseudo image to obtain a feature map of [64, 496, 432].

[0081] Step four: Send the pseudo image obtained in the previous step into the deep separable backbone network based on channel attention to extract features and obtain a 2D feature map.

[0082] Step four, the size of the input image is [64, 496, 432], which is down-sampled three times to save three feature maps of different scales, [64, 248, 216], [128, 124, 108], and [256, 62, 54], respectively. Then different transpose convolutions are used to restore the unified size [384, 124, 108], and finally the features containing different scales are obtained by concatenation.

[0083] Where the convolution of the network is all reduced by depth separable convolution to reduce the parameter amount of the model, as shown in Figure 4 The specific structure of the network is shown, wherein the Bneck structure also adopts a channel-based attention mechanism to improve the feature extraction capability of the model. And contains the SEBlock channel attention module, the main function of which is to let the network pay attention to those more important feature layers, as shown in Figure 5 The processing flow of Bneck and SEBlock is shown.

[0084] And a residual structure is also adopted to prevent the situation of gradient vanishing and gradient explosion, as shown in Figure 6 Through the residual link edge, the model can avoid the situation of gradient vanishing and gradient explosion.

[0085] Step five: send the 2D feature map into the target detection head based on SSD to get the result.

[0086] Specifically, step five will first pass the feature map obtained in step four through 3 different prediction heads, namely the classification head, the regression head, and the direction prediction head. Different prediction heads will give different prediction results. For the classification prediction head, the classification category is 3 categories, and the cross-entropy loss is calculated; for the regression prediction head, the position information of the 3D object is mainly predicted, and the coordinates of 7 objects are predicted; for the direction prediction head, the binary cross-entropy loss is calculated.

[0087] For the ground truth label, each object has 7 parameters, namely [x, y, z, w, l, h, θ], wherein the first three represent x, y, z in the laser radar coordinate system, w, l, h represent width, length, and height, respectively, and θ represents the heading angle.

[0088] The SmoothL1 regression loss is used between the real ground label and the prediction box, and the formula of the regression loss is:

[0089]

[0090] Δθ=sin(θ t -θ a )

[0091] Where x t y t z t represents the ground truth label, x a y a z a represents the prediction box, w t l t h t represents the width, length, and height of the label, w a la h a represent the predicted width, length, height. θ t θ a respectively represent the label turning angle and the predicted turning angle. The overall regression loss is:

[0092]

[0093] Considering that in a three-dimensional environment, the positioning loss cannot identify the angle of two completely opposite positioning frames, the softmax heading angle classification loss is used to learn the heading angle of the three-dimensional object. The object classification loss used is focalloss:

[0094] L cls = -α a (1-p a ) γ log p a

[0095] In the formula, the classification probability score is represented by p a , where the hyperparameters α a = 0.25, γ = 2, and the overall loss is represented as:

[0096]

[0097] where N pos represents the number of anchor boxes assigned as positive samples, and weights are again added to different losses because the 3D object detection task is more concerned about the coordinate information of the positioned object in the 3D world, so β loc = 2, β cls = 1, and β dir = 0.2.

[0098] The evaluation index used in step 5 is the interpolated AP|R40 of the KITTI standard to evaluate the performance of the model, as follows:

[0099]

[0100] where R = [0, 1 / 40, 2 / 40, …, 1] represents different recall rate thresholds, and P represents the accuracy under different R (recall rate). Different (IoU) thresholds (0.7, 0.5, 0.5) correspond to cars, pedestrians, and cyclists, respectively. If the defined IoU is lower, those predicted boxes are filtered out. Finally, non-maximum suppression (NMS) is used to retain the highest confidence prediction box.

[0101] Step six: mark the prediction box obtained in step five to the original point cloud, generate a new point cloud with detection results, visualize the prediction results, realize the 3D target detection task of the automatic driving car in the 3D world, and then the data can be transmitted to the downstream module and handed over to the decision module to make the next decision action according to the predicted 3D target.

[0102] Experimental results

[0103] Table 1, comparison results between different models

[0104]

[0105] According to the results in the above table 1, the model of the application greatly reduces the parameter amount required by the model under the premise of maintaining the accuracy, only needs 1M more parameters to achieve good mAP effect, greatly reduces the parameter amount of the network, realizes the light weight, and achieves good effect in the aspect of detecting vehicles.

[0106] In the application, the transfomer technology and the depth separable convolution technology are introduced in the feature extraction part, the Pillars Feature Net module in PointPillars is improved, the local transformer structure is added in the Pillars Feature Net module, the extraction ability of the network for the context semantic features can be improved, and the inaccuracy of the 3D target data detected by the laser radar is avoided; through the depth separable convolution technology, the parameter amount of the network can be greatly reduced, the light weight is realized, and the 3D target detection in the automatic driving car is realized.

[0107] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the features disclosed herein. The specification and examples given are considered exemplary only, and the true scope and spirit of the application is indicated by the claims.

Claims

1. A lightweight PointPillars autonomous driving 3D target detection method based on columnar self-attention, characterized in that, The method comprises the following steps: Step 1: obtaining point cloud data, image data and parameters of the device of a public road, wherein the parameters of the device include the internal and external parameters of the camera, the origin coordinates of the laser radar and the ground truth bounding box labeled; Step 2: cropping the point cloud data, and cropping the 360° point cloud around the vehicle body to only the point cloud appearing in the camera view; Step 3: then inputting the processed original point cloud into the first improved Pillars Feature Net module of PointPillars, adding a local self-attention transformer structure in the Pillars Feature Net module to improve the feature extraction capability, and finally obtaining a 2D pseudo image; Step 4: inputting the pseudo image obtained in the above step into a deep separable lightweight backbone network based on channel attention for feature extraction to obtain a 2D feature map; Step 5: inputting the 2D feature map into a target detection head based on SSD to obtain a detection result; Step 6: marking the detection result to the original point cloud to realize the 3D target detection task of the autonomous vehicle in the 3D world; The local self-attention transformer process of the Pillars Feature Net module in step 3 comprises the following steps: Step 3-1: Obtain the cropped raw point cloud data from step two, which contains all the points in the scene that appear in the camera's field of view , representing the three-dimensional world coordinates and surface reflection intensity of a single point, respectively; Step 3-2: point cloud columnization, inputting the obtained original point cloud data into a voxelize module for cutting, and calculating the corresponding coordinates of each column in the original three-dimensional world as a return value to the Pillars Feature Net module; Step 3-3: feature dimensionality increase, after obtaining the segmented column, calculating the cluster center of all points of each column, the center point offset of each column and the distance from the observer to the point, all of which are taken as the features of the point; Step 3-4: then inputting into the local self-attention transformer for local feature extraction and attention calculation; obtaining a tensor, wherein the first dimension represents how many columns a sample in the batch has, the second dimension represents how many points each column contains, and the third dimension represents the specific feature information of each point; inputting the tensor into the encoder structure of the transformer, and first performing position encoding on it; Then calculate the autocorrelation of each column; the final result sub-vector represents the importance of each point relative to other points; The multi-head attention mechanism is used to splice all the sub-vectors together, and then a weight to obtain the final attention result; then a residual structure and a forward propagation module are used for multiple iterations. Step 3-5: finally, the point cloud features after the encoder are restored to a 2D pseudo image according to the relative coordinates.

2. The cylinder self-attention-based lightweight PointPillars autonomous driving 3D target detection method according to claim 1, characterized in that, The method is characterized in that The 2D feature extraction network in step 4 comprises: After obtaining the 2D pseudo image through step 3, a lightweight network is redesigned. The size of the input image is Through three times of downsampling, three feature maps with different scales are saved Then, different transposed convolutions are performed to restore the unified size Finally, the feature maps containing different scale features are obtained by splicing. Among them, the convolution of the network adopts deep separable convolution to reduce the parameter quantity of the model, and the Bneck structure adopts a channel-based attention mechanism to improve the feature extraction capability of the model; and the Bneck structure contains an SEBlock channel attention module, which mainly functions to enable the network to focus on more important feature layers; and a residual structure is adopted to prevent gradient disappearance and gradient explosion.

3. The cylinder self-attention based lightweight PointPillars autonomous driving 3D target detection method according to claim 2, characterized in that, In step four, all parameters of the convolutional layer are dropped from down to ; The total parameter quantity is 24.3% of that of the original model.

4. The cylinder self-attention based lightweight PointPillars autonomous driving 3D target detection method according to claim 1, characterized in that, The SSD detection head in step five detects includes: Step 5-1: The 2D feature map obtained in step four is first subjected to three different prediction heads, namely a classification head, a regression head and a direction prediction head; different prediction results are obtained for different prediction heads; for the classification prediction head, the classification category is three, and a cross-entropy loss is calculated; for the regression prediction head, the position information of the 3D object is mainly predicted, specifically, the coordinate information of seven objects is predicted; for the direction prediction head, a binary cross-entropy loss is used for calculation; Step 5-2: For ground truth labels, each object has 7 parameters, respectively where the first three represent the width, length, and height in the laser-radar coordinate system, respectively , in order, and the last one represents the heading angle. ​ The SmoothL1 regression loss is used between the real ground mark and the prediction box, and the formula of the regression loss is: ， ， ； ， ， ； ； wherein represents a ground truth label, represents a predicted box, represents a width, length, height of a label, represents a predicted width, length, height; ; respectively represent a turning angle of a label and a predicted turning angle. The overall regression loss is: ； Considering that in a three-dimensional environment, the positioning loss cannot identify the angle of two completely opposite positioning frames, a softmax heading angle classification loss is used to learn the heading angle of the three-dimensional object; The object classification loss used is focal loss: ： where the classification probability score is denoted by where the hyperparameters = 0.25, = 2; The overall loss is represented as: ； wherein, represents the number of anchor boxes assigned as positive samples, and again weights are added for different losses, because more attention is paid to the coordinate information of the object positioned in the 3D world in 3D target detection, and thus , .

5. The cylinder self-attention based lightweight PointPillars autonomous driving 3D target detection method according to claim 4, characterized in that, The prediction results obtained in step 5-2 are evaluated: In general, the main performance parameter for 3D object recognition is the average precision, AP; it is slightly modified in the KITTI benchmark, using the interpolation of the KITTI standard to evaluate the performance of the model: ； wherein, represent different recall rate thresholds, represent the precision under different recall rates; different IoU thresholds (0.7, 0.5, 0.5) correspond to cars, pedestrians and cyclists respectively, and low-confidence prediction boxes are filtered out; finally, the non-maximum suppression NMS is adopted to retain the prediction box with the highest confidence.

6. The cylinder self-attention based lightweight PointPillars autonomous driving 3D target detection method according to claim 1, characterized in that, In step six, the detection results obtained in step five are marked into the original point cloud to generate a new point cloud with detection results, the prediction results are visualized, and the 3D target detection task of the autonomous vehicle in the 3D world is realized.

Images