Vehicle intelligent driving decision optimization method based on deep learning

Abstract

The invention discloses a vehicle intelligent driving decision optimization method based on deep learning. The method belongs to a vehicle intelligent driving decision technology. Firstly, various sensors are installed on a vehicle to collect data, and data sets are divided according to a proportion after cleaning and marking. A model containing an encoder-decoder and an attention mechanism is constructed, the weight is initialized, and a hyper-parameter and anti-overfitting strategy is set. And training the model by using a specific algorithm, calculating a loss value and evaluating and adjusting according to a verification set. After the model is integrated to the system, a preliminary candidate decision is determined according to a probability threshold, the decision is optimized in combination with a vehicle surrounding environment information set and a risk assessment function, and decision probabilities are redistributed in real time according to environment changes by means of an environment perception feedback mechanism. The method overcomes the defects of the traditional technology, utilizes the advantages of deep learning, effectively improves the driving decision precision, enhances the adaptability of the system to complex road conditions and environments, and guarantees the safe and efficient operation of intelligent driving.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle intelligent driving decision-making, and in particular to a vehicle intelligent driving decision-making optimization method based on deep learning. Background Art

[0002] With the rapid development of the automobile industry and the continuous advancement of artificial intelligence technology, intelligent driving has become a research hotspot and development trend in the field of transportation. Intelligent driving aims to use various advanced technologies to enable vehicles to automatically complete driving tasks without the continuous intervention of human drivers, improve traffic efficiency, reduce traffic accidents and provide passengers with a more convenient and comfortable travel experience.

[0003] However, the existing intelligent driving decision-making technology for vehicles still has many shortcomings. Traditional intelligent driving decision-making methods often rely on simple rules and limited sensor information processing methods. For example, in terms of perception of the vehicle's surrounding environment, analysis may only be based on a single type of sensor data, such as relying solely on millimeter-wave radar to determine the distance of the vehicle in front, but failing to fully integrate multi-source sensor data such as lidar and cameras, which makes the understanding of the surrounding environment not comprehensive and accurate enough. In complex traffic scenarios, such as intersections of urban roads, construction sections, or congested traffic on highways, it is difficult to accurately judge the true state and movement trend of surrounding vehicles, pedestrians, and obstacles by relying on a single sensor or simple rules, which can easily lead to decision-making errors, such as misjudging the timing of lane changes or failing to identify sudden pedestrians crossing the road in time, thereby causing traffic accidents or traffic jams, seriously affecting the safety and smoothness of driving.

[0004] In terms of model building and decision generation, some existing methods mostly use artificially designed feature rules to build decision models. This approach not only requires a lot of professional knowledge and experience, but also has difficulty adapting to complex and changing traffic environments and massive data features. Due to the limitations of human cognition, artificially designed feature rules may not cover all potential traffic conditions and environmental changes, resulting in poor generalization of the model when faced with new and unseen scenarios. For example, when encountering a special road sign or a new type of traffic participant (such as a special-purpose vehicle or a new type of intelligent vehicle), a model based on artificially designed feature rules may not be able to make the right decision, thus limiting the scope of application and reliability of the intelligent driving system.

[0005] In addition, the existing vehicle intelligent driving decision-making technology also has obvious deficiencies in model optimization and dynamic adjustment. In the process of model training, there is a lack of effective mechanisms to prevent the occurrence of overfitting and underfitting. Overfitting will cause the model to over-learn the noise and details in the training data, while ignoring the overall characteristics and laws of the data, resulting in reduced adaptability to new data in practical applications; underfitting means that the model has not fully learned the effective information in the data and cannot accurately predict driving decisions. At the same time, when the surrounding environment changes during vehicle driving, it is difficult for the existing decision-making system to dynamically optimize and adjust the decision in real time. For example, if you suddenly encounter bad weather or temporary road control during driving, the existing intelligent driving decision-making system may not be able to quickly adjust the driving strategy according to environmental changes, such as adjusting the speed, changing the driving route, etc., and thus cannot ensure driving safety and efficiency.

[0006] In summary, the existing vehicle intelligent driving decision-making technology has shortcomings in environmental perception, model construction and decision generation, model optimization and dynamic adjustment, and it is difficult to meet the increasingly complex traffic needs and safety standards. Summary of the invention

[0007] The main purpose of the present invention is to provide a vehicle intelligent driving decision optimization method based on deep learning, which can effectively solve the problems mentioned in the background technology.

[0008] To achieve the above object, the technical solution adopted by the present invention is: A vehicle intelligent driving decision optimization method based on deep learning comprises the following steps: S1. Data collection and compilation S1.1. Multi-source data collection: Install laser radar, high-definition camera, millimeter-wave radar and vehicle-mounted sensors on the vehicle to collect the vehicle's own status, surrounding environment information and driver operation data under different road conditions, environments and time periods, collect at the corresponding frequency and record the timestamp; S1.2, Data cleaning and labeling: Develop data cleaning algorithms, set abnormal data judgment rules for lidar, camera images, vehicle status and radar data, remove abnormal data, and professional labeling teams use professional software to label images with traffic signs, lane lines, vehicle and pedestrian information, and label radar data with obstacle categories and dynamic information based on image labeling results; S1.3, Data division: Divide the dataset into training set, validation set and test set in a ratio of 7:2:1, and ensure that the distribution of each scenario is similar; S2. Deep learning model construction S2.1. Model architecture design: Construct an encoder-decoder structure model; in the encoder, multiple CNN layers and pooling layers process image data in sequence, and the convolution kernel size, number, step size and pooling layer parameters of the CNN layer are set in sequence. After conversion by the fully connected layer, the vehicle state and radar data processed by the normalization are spliced; the fully connected layer and LSTM layer of the decoder process the spliced ​​data, and the number of hidden units and layers of the LSTM layer are set. The model introduces an attention mechanism, and the attention score, distribution and context vector are calculated and spliced ​​with the LSTM hidden state as the input of the output layer. The output layer outputs the probability of each driving decision through the fully connected layer and the softmax function; S2.2. Model initialization and hyperparameter setting: Use the Xavier initialization method to initialize the model weight matrix, and initialize the bias term to 0; set the learning rate, batch size, training cycle, and regularization parameters. The learning rate uses an exponential decay strategy, and the loss function adds a regularization term to prevent overfitting. S3. Model training and validation S3.1. Model training process: Use the training set data to train the model with the stochastic gradient descent combined with momentum optimization algorithm. Calculate the multi-classification cross entropy loss function value for each training batch of data. Regularly evaluate with the validation set to calculate the accuracy, recall, and F1 value. S3.2, Model verification and adjustment: Determine overfitting or underfitting based on the performance indicators of the validation set; in case of overfitting, increase the amount of training data, adjust the regularization parameters, and use dropout technology; in case of underfitting, increase the model complexity and adjust the hyperparameters; use visualization tools to monitor the loss value curve, accuracy curve, and parameter distribution during the training process; S4, Intelligent Driving Decision Optimization S4.1. Decision optimization strategy: Integrate the trained model into the intelligent driving system. The model receives sensor data and outputs the probability distribution of driving decisions. Set the probability threshold to determine the preliminary candidate decisions, build the vehicle's surrounding environment information set and risk assessment function, and obtain the candidate decision risk value by integrating the environmental information elements. Adjust the driving decision probability according to the risk adjustment coefficient and the specific calculation formula to obtain the optimized decision. Introduce the environmental perception feedback mechanism. When the vehicle makes a decision, the environmental perception system monitors the environmental changes and builds the information set. Through the feedback adjustment function, correlation coefficient and calculation formula, the driving decision probability is redistributed based on the environmental changes to adapt to the dynamic changes of the environment.

[0009] Preferably, in S1.1, the laser radar scanning frequency reaches 15Hz, the camera resolution is 1920×1080 pixels, the frame rate is 30fps, the millimeter wave radar distance detection accuracy is 0.05 meters, the speed is 0.05m / s, and the angle is 0.05 degrees. The time synchronization technology based on the atomic clock is used to connect the sensors with the help of the dedicated data bus inside the vehicle, so that the data time synchronization error is controlled within 1 microsecond; Arrange test vehicles to drive on urban roads (covering busy sections of the city center, ordinary streets, roads around residential areas, etc.), highways, and rural roads. Drive 150 kilometers of various types of urban roads; drive 300 kilometers of highways in different speed limit sections and traffic flow periods; drive a total of 200 kilometers of rural roads on unpaved and paved roads of different terrains. Collect data in sunny, rainy, snowy, foggy weather and during morning rush hour, noon, evening rush hour, and late night. The vehicle's own status data is collected at a frequency of 120Hz, environmental data is collected at the frequency of each sensor, and driver operation data and vehicle status data are collected synchronously and recorded with timestamps.

[0010] Preferably, in S1.2: Data cleaning: LiDAR data: Point cloud distance values ​​less than 0.05 meters or greater than 600 meters will be deleted; point cloud density in a certain area with abnormal change rate within 5 consecutive scanning cycles (determined by the vehicle motion model) will be marked for review. Camera image data: Low image clarity index (threshold determined by the image gradient algorithm), brightness mean value exceeding the normal range by 3 times the standard deviation or contrast lower than 0.15 will be discarded. Vehicle status and radar data: Data such as speed and acceleration will be deleted if they exceed the mean ±4 times the standard deviation; Data annotation: 60 common categories of traffic signs are classified and annotated, and the boundaries are accurately annotated with polygons; lane lines are drawn and annotated according to the center line and 4 types are annotated; vehicles and pedestrians are accurately positioned with rectangular boxes (accurate to pixels) and 6 directions of movement are annotated. Radar data annotation: Combine image annotation with radar features to annotate 8 common obstacle categories and dynamic information (moving or not, speed and direction accurate to 0.05m / s and 0.05 degrees). The annotations are reviewed twice to ensure that the error rate is less than 0.5%, and are stored in a standardized format (such as PASCAL VOC format XML files).

[0011] Preferably, all data in S1.3 are classified according to road conditions (city, highway, rural), environment (4 types of weather, 4 time periods), and driving scenarios. The data in each subset are numbered and stratified sampled according to the ratio of 7:2:1. First extract data to the training set, then sample the validation set from the remaining, and the rest is the test set. For example, urban road data accounts for 38% of the training set, and data on busy sections in the city center accounts for 28% of the urban road training data; urban road data in the validation set and test set account for approximately 10% and 2% respectively, and the proportion of data in different scenarios in each subset is similar to that in the training set, with an error within 3%. After division, they are stored in independent databases to record data sources and attribute information.

[0012] Preferably, the model architecture design method in S2.1 includes: (1) Encoder part: ①CNN layer construction Construct three sequentially connected CNN layers to process image data; The first layer of CNN, convolution kernel 3×3, number 32, stride 1 and zero padding, input image , after convolution and bias and ReLU activation, the output is , and its calculation formula is: ,in represents the convolution operation, is the bias term; The second layer of CNN uses 5×5 convolution kernels, the number is 64, the stride is 2, no padding, and the input , output ; ; The third layer of CNN uses 7×7 convolution kernel, number 128, stride 2, no padding, input , output , ; ②Pooling layer settings Connect two maximum pooling layers, the pooling kernel is 2×2, the stride is 2, and the first pooling layer input is have to : ; The second pooling layer input have to : ; ③Fully connected layer integration The last fully connected layer converts the output of the pooling layer into a suitable dimension for concatenation with other data. Let the output dimension be , the input is , the weight matrix is , the bias is , then the output ; (2) Data fusion and decoder input preparation: The normalized vehicle status data , the dimension is , and radar data , the dimension is , and the encoder’s fully connected layer output After splicing, the dimension of the vector after splicing is , denoted as ; (3) Decoder part ①Fully connected layer feature fusion The first fully connected layer, weight , bias ,enter , output dimension , activated by ReLU: ; The second fully connected layer has a weight of W d2 , bias b d2 , output dimension : ; ②LSTM layer timing learning Let the first LSTM layer input , the number of hidden units , the relevant weight matrix , , , , bias , , , , the current input is , the hidden state at the previous moment is , the cell state at the last moment is , the Forget Gate , input gate , candidate cell states , cell state update , output gate , hide status updates ; The second LSTM layer is similar, with the number of hidden units set to ; ③Attention mechanism Set the decoder LSTM layer hidden state , encoder output , attention score , the attention distribution is obtained by softmax normalization , the context vector , concatenated with the LSTM hidden state as the output layer input; (4) Output layer: Set weight , bias , the input is the concatenation vector of attention and LSTM output, and the output node corresponds to the number of driving decision categories , the calculation formula is , and the decision probability is obtained by softmax ,in For the The probability of a driving decision.

[0013] Preferably, the S2.2 model initialization and hyperparameter setting specifically includes: (1) Model initialization: Use Xavier to initialize weights , the formula is , bias Initialized to 0; (2) Hyperparameter setting: The initial value of the learning rate is 0.001, the batch size is 32, the training cycle is 100, the regularization parameter is 0.0001 and L2 regularization is used, and the loss function , =0.0001, is the cross entropy loss.

[0014] Preferably, the S3.1 model training process is as follows: (1) Optimization algorithm settings: The model is trained using an optimization algorithm combining stochastic gradient descent and momentum, and the momentum coefficient is set to , let the model parameters be , the learning rate is , the current batch data gradient is , the parameter update formula is ,in is the last parameter update amount, the new parameter , learning rate The initial value is set to 0.001, and the exponential decay learning rate strategy is adopted. The formula is: ,in , every time you pass training cycles, the learning rate decays to the original times; (2) Loss function calculation: Using multi-classification cross entropy loss function , let the model prediction output be (Probability distribution of each driving decision after processing by the softmax function), the true label is (One-hot encoding form), the number of samples is , whose formula is ,in is the number of driving decision categories. After each training batch of data, the batch loss value is calculated according to this formula; (3) Regular evaluation operations: Every 5 training cycles, the model is evaluated using the validation set data and the accuracy of the model on the validation set is calculated. , recall rate , Value; Accuracy calculation formula ,in represents true positives (the number of samples predicted by the model to be positive and actually positive), represents true negative examples (the number of samples that the model predicts to be negative and are actually negative), represents false positives (number of samples predicted by the model to be positive but actually negative), represents false negative examples (the number of samples predicted by the model to be negative but actually positive); the recall rate calculation formula ; Value calculation formula .

[0015] Preferably, the specific steps of S3.2 model verification and adjustment are as follows: (1) Overfitting and underfitting discrimination: If the model's loss value in the training set continues to decrease, while the performance index of the validation set no longer improves or even decreases, it is considered to be overfitting; if the performance index of the validation set is always low and the loss value decreases slowly, it is considered to be underfitting; (2) Adjustment strategy implementation: When overfitting: Increase the amount of training data and collect more vehicle driving data; Adjust the regularization parameters, such as Regularization parameter Increasing from 0.0001 to 0.001, the loss function becomes ; Use dropout technology to set the dropout probability in the fully connected layer or LSTM layer ; When underfitting: Increase the model complexity, increase the number of neurons or add network layers in the fully connected layer or LSTM layer of the decoder, and pay close attention to the performance of the validation set after adjustment; Adjust the hyperparameters and set the learning rate Increase from 0.001 to 0.01 and increase the training cycle while paying attention to preventing overfitting; (3) Application of visual monitoring: Use the TensorBoard visualization tool to monitor the training process, draw the loss value curve, observe the changes in the loss values ​​of the training set and the validation set with the training cycle, and judge overfitting or underfitting based on this; draw the accuracy curve and parameter distribution curve, the former shows the model learning effect, and the latter reflects the changes in model parameters.

[0016] Preferably, the decision optimization strategy in S4.1 is specifically: (1) Model integration and preliminary decision-making The deep learning model is integrated into the intelligent driving system, connected to the vehicle sensor data interface, and receives multi-source data; the model processes the data according to its neural network architecture, outputs the probability values ​​corresponding to multiple driving decisions, and forms a driving decision probability distribution vector ,in is the number of driving decision categories and satisfies ,in Indicates The probability of a driving decision, setting the probability threshold , ,when , corresponding to the decision Become a preliminary candidate decision ; (2) Risk assessment and decision optimization Constructing vehicle surroundings information set ,in Design risk assessment functions for environmental information elements determined based on vehicle dynamics, traffic rules and safe driving criteria , which is based on the environment information set The elements in the formula are constructed, and preliminary candidate decisions are obtained through comprehensive operations on each element. The risk value under the current environment; the risk assessment function has the following form: ,in Based on The weighting factor for determining the importance in the risk assessment, It is for Constructed risk assessment subfunction; based on the risk assessment results, introduce the risk adjustment coefficient , value range , through the formula The probability of each driving decision is adjusted, where After adjustment driving decision probability values, and , thus obtaining optimized driving decisions; (3) Environmental perception feedback mechanism When the vehicle makes a driving decision When the vehicle is in the environment, the environmental perception system monitors the changes in the vehicle's surrounding environment in real time and builds an environmental change information set. ,in For environmental information elements Corresponding change; construct feedback adjustment function This function is constructed based on the inherent logical relationship between the vehicle kinematic model and environmental changes. When the environmental change meets certain conditions or is within a certain range, the formula The probability of each driving decision is adjusted, where It is the adjustment coefficient related to environmental changes, and its value depends on environmental factors. It is based on the feedback adjustment function A function is constructed to calculate the decision adjustment weights, so as to achieve real-time redistribution of driving decision probabilities according to environmental changes, so that driving decisions can adapt to dynamic changes in the environment.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. The deep learning model processes multi-source data, automatically learns the features and patterns therein, without manually setting feature rules, accurately predicts driving decisions, reduces deviations, improves the accuracy of predicting various driving behaviors in complex traffic scenarios, reduces errors, and ensures safe and smooth driving.

[0018] 2. During training, the system is exposed to multiple road conditions, environments and scenario data, and the correlation patterns are explored. When encountering complex and diverse actual scenarios, the system can quickly adapt and make reasonable decisions, cope with urban traffic flow and bad weather conditions, and enhance the practicality and reliability of the system.

[0019] 3. Use Xavier initialization, reasonable hyperparameters and L2 regularization to control model complexity and prevent overfitting. Use validation set evaluation and visual monitoring to promptly detect and adjust underfitting, so as to balance model training and generalization and ensure stable and accurate operation of the system.

[0020] 4. After being integrated into the system, the initial decision is made based on the probability threshold and risk assessment, and the probability distribution of multiple factors is considered. With the help of the environmental perception feedback mechanism, the driving decision probability is adjusted according to the environmental change information through specific calculations to respond to emergencies and changes in road conditions, improve the system's response and adaptability, and improve driving safety and traffic efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of the vehicle intelligent driving decision optimization method based on deep learning of the present invention. DETAILED DESCRIPTION

[0022] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0023] like Figure 1 The process steps shown are further described below in conjunction with a specific implementation case to illustrate the present invention: 1. Data collection and organization Multi-source data collection Advanced laser radar, high-definition camera, millimeter-wave radar and vehicle-mounted sensors are installed on a test vehicle with intelligent driving function development. Among them, the laser radar scanning frequency is precisely set to 15Hz to ensure that the dynamic changes of the vehicle's surrounding environment can be captured in time; the camera resolution is as high as 1920×1080 pixels and the frame rate is 30fps, which can clearly capture the image information around the vehicle; the millimeter-wave radar has a distance detection accuracy of 0.05 meters, a speed accuracy of 0.05m / s, and an angle accuracy of 0.05 degrees, which can accurately detect the position and movement status of surrounding vehicles and obstacles.

[0024] By using high-precision time synchronization technology based on atomic clocks and connecting various sensors with the help of a dedicated data bus inside the vehicle, the data time synchronization error is strictly controlled within 1 microsecond, ensuring the time consistency and accuracy of the collected data.

[0025] The test vehicles were arranged to collect driving data under different road conditions such as urban roads (including busy sections in the city center such as commercial streets, ordinary streets such as streets near residential areas, and roads around residential areas, etc.), highways, and rural roads. On urban roads, the mileage of each category reached 150 kilometers. For example, in busy sections in the city center, vehicles frequently start and stop, and there are dense vehicles and pedestrians around, so data under complex traffic conditions can be collected; on highways, they were driven 300 kilometers in different speed limit sections (such as sections with speed limits of 60km / h, 100km / h, 120km / h, etc.) and traffic flow periods (time periods with large traffic volume during the morning rush hour, relatively small traffic volume at noon, and increased traffic volume during the evening rush hour), so data under different driving speeds and traffic flows can be obtained; on rural roads, they were driven 200 kilometers in total on non-paved roads (such as muddy roads and gravel roads) and paved roads of different terrains, so data under different road conditions can be collected.

[0026] Data is collected in different weather conditions such as sunny, rainy, snowy, and foggy, and at different time periods such as morning rush hour, noon, evening rush hour, and late at night, covering all possible driving scenarios. The vehicle's own status data (such as speed, acceleration, steering angle, etc.) is collected at a frequency of 120Hz, environmental data is collected at the frequency of each sensor, and driver operation data (such as braking, accelerator, shifting, etc.) is collected synchronously with the vehicle status data and accurately recorded with timestamps, providing rich and accurate raw data for subsequent data analysis.

[0027] Data cleaning and annotation Data cleaning For lidar data, if the point cloud distance value is less than 0.05 meters, it may be sensor noise or close-range misjudgment. If it is greater than 600 meters, it exceeds the effective detection range and the data will be deleted. If the point cloud density in a certain area has an abnormal change rate within 5 consecutive scanning cycles (for example, according to the vehicle motion model, the surrounding point cloud density should change relatively smoothly during normal driving. If there is a sudden and large fluctuation, there may be a problem), it will be marked for review to further confirm the reliability of the data.

[0028] For camera image data, the image clarity index is determined by the image gradient algorithm. If the image clarity is low, the brightness mean exceeds the normal range by 3 standard deviations (for example, under normal circumstances, the image brightness should be within a certain range. If it exceeds this range, it may be caused by excessive light or sensor failure), or the contrast is lower than 0.15, the image data will be discarded to ensure data quality.

[0029] In the vehicle status and radar data, if the speed, acceleration and other data exceed the mean ±4 times the standard deviation, it indicates that there may be data anomalies and they should be deleted to avoid interference with subsequent model training.

[0030] Data Annotation The professional annotation team uses professional annotation software to carefully annotate the images. For traffic signs, 60 common categories are classified and annotated, such as prohibition signs, instruction signs, warning signs, etc., and the boundaries are accurately annotated with polygons to ensure the accuracy of the annotation; lane lines are drawn according to the center line and annotated in 4 types, including single solid line, double solid line, single dashed line, double dashed line, etc.; vehicles and pedestrians are accurately positioned with rectangular boxes (accurate to pixels) and 6 movement directions are annotated, such as forward, backward, left turn, right turn, stationary, acceleration or deceleration, etc.

[0031] When labeling radar data, we combine image labeling with radar features to label 8 common obstacle categories and dynamic information, such as cars, motorcycles, pedestrians, bicycles, large trucks, road construction facilities, traffic cones, etc., and accurately label whether they are moving, speed and direction to 0.05m / s and 0.05 degrees. After labeling, we conduct two rounds of strict review to ensure that the error rate is less than 0.5%, and finally store them in a standardized format (such as PASCAL VOC format XML file) to facilitate subsequent data processing and model training.

[0032] Data partitioning First, all data are classified in detail according to road conditions (city, highway, rural), environment (4 types of weather, 4 time periods), and driving scenarios. The data are numbered in each subset, and then stratified sampling is performed according to the ratio of 7:2:1. For example, urban road data accounts for 38% of the training set, of which the data of busy sections in the city center accounts for 28% of the urban road training data; urban road data in the validation set and test set account for about 10% and 2%, and the proportion of data in different scenarios in each subset is similar to that in the training set, with an error of less than 3%. First extract data to the training set, then extract the validation set from the remaining data, and finally the remaining is the test set. After division, they are stored in independent databases, and the data source and attribute information are recorded in detail to provide orderly data support for subsequent model training and evaluation.

[0033] 2. Deep Learning Model Construction Model architecture design Encoder part CNN layer construction Construct three CNN layers connected in sequence to process image data. The first layer of CNN has a convolution kernel of 3×3, a number of 32, a stride of 1 and zero padding. Take an input image For example, during the convolution process, each 3×3 convolution kernel slides on the image, extracts features from the local area of ​​the image, and obtains the output after convolution, bias and ReLU activation. , and its calculation formula is: The ReLU activation function here plays the role of introducing nonlinear factors to enhance the expressiveness of the model. For example, when the feature value after convolution is negative, the ReLU function will set it to 0, avoiding the gradient vanishing problem and enabling the model to learn more complex image features.

[0034] The second layer of CNN uses 5×5 convolution kernels, the number is 64, the stride is 2, and there is no padding. , further expand the receptive field through larger convolution kernels and strides, extract more abstract features, and output : .

[0035] The third layer of CNN uses 7×7 convolution kernel, number 128, stride 2, no padding, input , output : Through the progressive construction of these three CNN layers, valuable feature information for vehicle intelligent driving decision-making is gradually extracted from the image.

[0036] Pooling layer settings Connect two maximum pooling layers, with pooling kernels of 2×2 and stride 2. The first pooling layer input have to : Through the maximum pooling operation, the data dimension is reduced without changing the number of features, the amount of subsequent calculations is reduced, and the main feature information is retained. For example, in a 4×4 feature map, after the pooling operation of the 2×2 pooling kernel, it becomes a 2×2 feature map, and the maximum value in each pooling kernel area is selected as the output.

[0037] The second pooling layer input have to : , further compressing the data dimension.

[0038] Fully connected layer integration The last fully connected layer converts the output of the pooling layer into a suitable dimension for concatenation with other data. Let the output dimension be , the input is , the weight matrix is , the bias is , then the output , converting image features into a form suitable for fusion with other data.

[0039] Data fusion and decoder input preparation The normalized vehicle status data (Dimensions are , such as vehicle speed, acceleration, steering angle and other data after normalization) and radar data (Dimensions are , such as the distance and speed of surrounding vehicles detected by radar) and the fully connected layer output of the encoder The vector dimension after splicing is , denoted as , realizing the fusion of multi-source data and providing more comprehensive input information for the decoder.

[0040] Decoder part Fully connected layer feature fusion The first fully connected layer, weight , bias ,enter , output dimension , activated by ReLU: , further feature fusion and nonlinear transformation are performed on the input data.

[0041] The second fully connected layer has a weight of W d2 , bias b d2 , output dimension : , further adjust the data characteristics.

[0042] LSTM layer temporal learning Let the first LSTM layer input , the number of hidden units , the relevant weight matrix , , , Bias , , , , the current input is , the hidden state at the previous moment is , the cell state at the last moment is , the Forget Gate , which is used to determine which information of the cell state at the previous moment is retained; the input gate , used to control the input of new information; candidate cell states , generate candidate cell states at the current moment; cell state update , update the cell state according to the results of the forget gate and the input gate; the output gate , determine the hidden state information of the output; hidden state update For example, during vehicle driving, the LSTM layer can learn the vehicle's movement trends and driving rules based on previous driving status and current input information, such as in the case of continuous acceleration or deceleration, and can better predict subsequent driving decisions.

[0043] The second LSTM layer is similar, with the number of hidden units set to , further process time series data and enhance the model’s learning ability for time series information.

[0044] Attention Mechanism Set the decoder LSTM layer hidden state , encoder output . Attention score , by calculating the correlation score between the current hidden state of the decoder and the output of the encoder, and normalizing it with softmax to get the attention distribution , determines the importance weight of each encoder output at the current moment. The context vector , extract the important information related to the current moment from the encoder output and concatenate it with the LSTM hidden state as the output layer input, so that the model can better focus on important input information and improve the accuracy of decision-making.

[0045] Output Layer Set weight , bias , the input is the concatenation vector of attention and LSTM output, and the output node corresponds to the number of driving decision categories (e.g. acceleration, deceleration, constant speed, left turn, right turn, lane change, etc.), the calculation formula is , and the decision probability is obtained by softmax , and finally outputs the probability distribution of various driving decisions of the vehicle in the current situation.

[0046] Model initialization and hyperparameter setting Model initialization: Use Xavier to initialize weights , the formula is This initialization method can make the parameter distribution of the model more reasonable in the early stage of training, which helps to speed up the training and improve the model performance. Initialized to 0.

[0047] Hyperparameter settings: initial learning rate value 0.001, batch size 32, training cycle 100, regularization parameter 0.0001 and L2 regularization, loss function ( =0.0001, The learning rate adopts an exponential decay strategy, which gradually decreases with the increase of training cycles. For example, after every 10 training cycles, the learning rate decays to 0.9 times of the original value, so that the model can quickly learn data features in the early stage of training, and can adjust parameters more finely in the later stage to avoid overfitting. L2 regularization constrains parameters to prevent model parameters from being too large and improve the generalization ability of the model.

[0048] 3. Model training and verification Model training process Optimization algorithm setting: The model is trained using an optimization algorithm combining stochastic gradient descent and momentum, and the momentum coefficient is set to . Let the model parameters be , the learning rate is , the current batch data gradient is , the parameter update formula is ,in is the last parameter update amount, the new parameter For example, in a training batch, after calculating the gradient of the current batch of data, the model parameters are updated according to the momentum coefficient and learning rate, so that the model is optimized in the direction of reducing the loss function. The initial value is set to 0.001, and the exponential decay learning rate strategy is adopted. The formula is: ( ), after every 10 training cycles, the learning rate decays to 0.9 times of the original value.

[0049] Loss function calculation: using multi-classification cross entropy loss function . Let the model prediction output be (Probability distribution of each driving decision after processing by the softmax function), the true label is (One-hot encoding form), the number of samples is , whose formula is ,in is the number of driving decision categories. After each batch of data is trained, the batch loss value is calculated according to this formula to measure the prediction error of the model on the current batch of data and provide a basis for parameter updating.

[0050] Regular evaluation operation: every 5 training cycles, use the validation set data to evaluate the model and calculate the accuracy of the model on the validation set , recall rate , Value and other indicators. Accuracy calculation formula ,in represents true positives (the number of samples predicted by the model to be positive and actually positive), represents true negative examples (the number of samples that the model predicts to be negative and are actually negative), represents false positives (number of samples predicted by the model to be positive but actually negative), represents false negative examples (the number of samples predicted by the model to be negative but actually positive); the recall rate calculation formula ; Value calculation formula These indicators are used to comprehensively evaluate the performance of the model on the validation set and understand the learning effect and generalization ability of the model.

[0051] Model validation and tuning Overfitting and underfitting discrimination: If the loss value of the model in the training set continues to decrease, and the performance indicators of the validation set (such as accuracy, recall, F1 value, etc.) no longer improve or even decrease, it is judged to be overfitting. For example, in the early stage of training, as the training cycle increases, the loss value of the training set gradually decreases from 0.8 to 0.4, but the accuracy of the validation set begins to decrease after reaching 80%, which indicates that the model may be overfitting. At this time, the model may have over-learned the noise and details in the training data, and the generalization ability of new data has deteriorated. If the performance indicators of the validation set are always low and the loss value decreases slowly, it is judged to be underfitting. For example, after multiple training cycles, the accuracy of the validation set has been hovering around 60%, and the loss value of the training set has not decreased significantly, which means that the model has not fully learned the characteristics and laws in the data and cannot accurately predict driving decisions.

[0052] Adjust strategy implementation When overfitting Increase the amount of training data, for example, arrange test vehicles to collect data in more different urban roads, highways and rural roads, or collect driving data in different seasons and time periods, so that the model can learn a wider range of scenario characteristics and reduce excessive reliance on specific data.

[0053] Adjust the regularization parameters, such as Regularization parameter Increasing from 0.0001 to 0.001, the loss function becomes , by increasing the weight of the regularization term, the model parameters are more strongly constrained to prevent the parameters from being too large, thereby reducing the degree of overfitting.

[0054] Use dropout technology to set the dropout probability in the fully connected layer or LSTM layer , for example, setting , randomly discard some neuron connections during the training process, so that the model has a certain degree of randomness in each training, reducing the collaborative adaptability between neurons and enhancing the generalization ability of the model.

[0055] When underfitting Increase the complexity of the model, increase the number of neurons in the fully connected layer or LSTM layer of the decoder, such as increasing the number of neurons in a fully connected layer from 128 to 256, or add additional network layers, such as adding another LSTM layer to the decoder part. After adjustment, pay close attention to the performance of the validation set to observe whether the model can better learn the data features.

[0056] Adjust the hyperparameters and increase the learning rate from 0.001 to 0.01, so that the model can learn data features faster during training. At the same time, increase the training cycle, for example, from 100 cycles to 150 cycles, but be careful to prevent overfitting. Regularly evaluate the performance of the model on the validation set while increasing the training cycle.

[0057] Visual monitoring application: Use the TensorBoard visualization tool to monitor the training process. Draw the loss value curve to observe the changes in the loss values ​​of the training set and the validation set as the training cycle progresses. If the loss value of the training set continues to decrease while the loss value of the validation set increases in the later stage, it can be intuitively judged that overfitting has occurred; if the loss values ​​of both decrease slowly, there may be underfitting. Draw the accuracy curve to clearly see the improvement trend of the model's accuracy during the training process. If the accuracy stagnates or decreases on the validation set, it also indicates that there may be a problem. Draw the parameter distribution curve to understand the changes in the model parameters during the training process, such as whether there are abnormal situations where the parameter values ​​are too large or too small, so as to promptly discover problems and adjust the model's training strategy to ensure that the model can be trained and optimized stably and efficiently.

[0058] 4. Intelligent Driving Decision Optimization Decision optimization strategy Model integration and preliminary decision-making: Integrate the trained deep learning model into the intelligent driving system and closely connect it to the vehicle sensor data interface to ensure that it can receive multi-source data in real time. For example, when the vehicle is driving, sensors such as lidar, cameras, and millimeter-wave radar continuously collect data and transmit it to the model. The model quickly processes this data based on its neural network architecture and outputs the probability values ​​corresponding to multiple driving decisions to form a driving decision probability distribution vector. ,in is the number of driving decision categories and satisfies . Set the probability threshold ,when When the corresponding decision Become a preliminary candidate decision For example, the driving decision probabilities output by the model are acceleration 0.2, deceleration 0.1, constant speed 0.1, left turn 0.7, right turn 0.05, and lane change 0.05. Since the probability of turning left is greater than 0.6, turning left becomes the preliminary candidate decision.

[0059] Risk assessment and decision optimization: Building a vehicle surrounding environment information set ,in It is an environmental information element determined based on vehicle dynamics, traffic rules and safe driving principles. For example, It can be the speed and distance information of surrounding vehicles. It is the curvature and slope information of the road, etc. Design risk assessment function , assuming that for the lane change decision, if there is a vehicle approaching quickly in the adjacent lane and the distance is relatively close (assuming the safety distance threshold is 5 meters), then the corresponding risk assessment subfunction The value increases, and the weight coefficient is determined according to its importance in risk assessment Based on the risk assessment results, the risk adjustment factor is introduced , through the formula Adjust the probability of each driving decision. If the risk assessment result is high, the corresponding probability is reduced. , and recalculate the probabilities of other decisions so that the sum of all decision probabilities is still 1, thus obtaining the optimized driving decision.

[0060] Environmental perception feedback mechanism: When the vehicle executes a driving decision When the vehicle is in the environment, the environmental perception system monitors the changes in the vehicle's surrounding environment in real time and builds an environmental change information set. ,in For environmental information elements Corresponding changes. For example, after the vehicle makes an acceleration decision, the relative speed and distance of the surrounding vehicles change, and these changes are captured and recorded by the environmental perception system. Construct feedback adjustment function , which is constructed based on the inherent logical relationship between the vehicle kinematic model and environmental changes. When the environmental change meets certain conditions or is within a certain range, the formula The probability of each driving decision is adjusted, where is the adjustment coefficient related to environmental changes, and its value depends on environmental factors. Assuming that an obstacle suddenly appears in front of the vehicle during acceleration, the environmental perception system detects this change and adjusts the driving decision probability according to the feedback adjustment function, so that the vehicle can make decisions such as braking or avoidance in time, thereby achieving real-time redistribution of driving decision probabilities according to environmental changes, making driving decisions adapt to dynamic changes in the environment and ensuring driving safety.

[0061] Through the above detailed embodiments, the specific operation process and application effect of the vehicle intelligent driving decision optimization method based on deep learning are fully demonstrated, which can effectively solve the shortcomings of the existing vehicle intelligent driving decision-making technology and improve the performance and reliability of the intelligent driving system.

[0062] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A vehicle intelligent driving decision optimization method based on deep learning, characterized in that: The following steps are involved: S1. Data collection and compilation S1.

1. Multi-source data collection: Install laser radar, high-definition camera, millimeter-wave radar and vehicle-mounted sensors on the vehicle to collect the vehicle's own status, surrounding environment information and driver operation data under different road conditions, environments and time periods, collect at the corresponding frequency and record the timestamp; S1.2, Data cleaning and labeling: Develop data cleaning algorithms, set abnormal data judgment rules for lidar, camera images, vehicle status and radar data, remove abnormal data, and professional labeling teams use professional software to label images with traffic signs, lane lines, vehicle and pedestrian information, and label radar data with obstacle categories and dynamic information based on image labeling results; S1.3, Data division: Divide the dataset into training set, validation set and test set in a ratio of 7:2:1, and ensure that the distribution of each scenario is similar; S2. Deep learning model construction S2.

1. Model architecture design: Construct an encoder-decoder structure model; in the encoder, multiple CNN layers and pooling layers process image data in sequence, and the convolution kernel size, number, step size and pooling layer parameters of the CNN layer are set in sequence, and the vehicle status and radar data processed after conversion by the fully connected layer are spliced ​​with the normalized data; The decoder's fully connected layer and LSTM layer process the spliced ​​data. The LSTM layer sets the number of hidden units and layers. The model introduces an attention mechanism, calculates the attention score, distribution, and context vector, and splices them with the LSTM hidden state as the output layer input. The output layer outputs the probability of each driving decision through the fully connected layer and the softmax function. S2.

2. Model initialization and hyperparameter setting: Use the Xavier initialization method to initialize the model weight matrix, and initialize the bias term to 0; set the learning rate, batch size, training cycle, and regularization parameters. The learning rate uses an exponential decay strategy, and the loss function adds a regularization term to prevent overfitting. S3. Model training and validation S3.

1. Model training process: Use the training set data to train the model with the stochastic gradient descent combined with momentum optimization algorithm. Calculate the multi-classification cross entropy loss function value for each training batch of data. Regularly evaluate with the validation set to calculate the accuracy, recall, and F1 value. S3.2, Model verification and adjustment: Determine overfitting or underfitting based on the performance indicators of the verification set; in case of overfitting, increase the amount of training data, adjust regularization parameters, and use dropout technology; in case of underfitting, increase model complexity and adjust hyperparameters; Use visualization tools to monitor loss value curves, accuracy curves, and parameter distribution during the training process; S4, Intelligent Driving Decision Optimization S4.1, Decision optimization strategy: Integrate the trained model into the intelligent driving system, and the model receives sensor data and outputs the driving decision probability distribution; A probability threshold is set to determine preliminary candidate decisions, and a vehicle surrounding environment information set and risk assessment function are constructed. The risk value of the candidate decision is obtained by integrating environmental information elements. The driving decision probability is adjusted according to the risk adjustment coefficient and a specific calculation formula to obtain an optimized decision. An environmental perception feedback mechanism is introduced. When the vehicle makes a decision, the environmental perception system monitors environmental changes and constructs an information set. Through feedback adjustment functions, correlation coefficients and calculation formulas, the driving decision probability is redistributed based on environmental changes to adapt to dynamic environmental changes.

2. The vehicle intelligent driving decision optimization method based on deep learning according to claim 1, characterized in that: The laser radar scanning frequency in S1.1 reaches 15Hz, the camera resolution is 1920×1080 pixels, the frame rate is 30fps, the millimeter wave radar distance detection accuracy is 0.05 meters, the speed is 0.05m / s, and the angle is 0.05 degrees. The time synchronization technology based on atomic clocks is used to connect sensors with the help of the dedicated data bus inside the vehicle, so that the data time synchronization error is controlled within 1 microsecond; The test vehicles were arranged to drive on urban roads, highways, and rural roads. The test vehicles drove 150 kilometers of various types of urban roads; 300 kilometers of highways in different speed limit sections and traffic flow periods; and 200 kilometers of rural roads on unpaved and paved roads in different terrains. Data was collected in sunny, rainy, snowy, and foggy weather, as well as during the morning rush hour, noon, evening rush hour, and late night. The vehicle's own status data was collected at a frequency of 120Hz, and environmental data was collected according to the frequency of each sensor. The driver's operation data and vehicle status data were collected synchronously and the timestamps were recorded.

3. The vehicle intelligent driving decision optimization method based on deep learning according to claim 2 is characterized in that: In S1.2: Data cleaning: LiDAR data: Point cloud distance values ​​less than 0.05 meters or greater than 600 meters will be deleted; point cloud density in a certain area with abnormal change rate within 5 consecutive scanning cycles will be marked for review; camera image data: images with low clarity index, brightness mean value exceeding the normal range by 3 times the standard deviation or contrast lower than 0.15 will be discarded; vehicle status and radar data: speed and acceleration data exceeding the mean ±4 times the standard deviation will be deleted; Data labeling: Classify and label traffic signs into 60 common categories and accurately mark the boundaries with polygons; draw lane lines according to the center line and label 4 types; use rectangular boxes to accurately locate vehicles and pedestrians and label 6 directions of movement; radar data labeling: combine image labeling with radar features to label 8 common obstacle categories and dynamic information; complete two rounds of review to ensure that the error rate is less than 0.5%, and store in a standardized format.

4. The vehicle intelligent driving decision optimization method based on deep learning according to claim 3 is characterized by: In S1.3, all data are classified according to road conditions, environment, and driving scenarios; the data in each subset are numbered and stratified sampling is performed according to the ratio of 7:2:1; data is first extracted to the training set, and then the validation set is sampled from the remaining data, and the rest is the test set; after division, they are stored in independent databases respectively, and the data source and attribute information are recorded.

5. The vehicle intelligent driving decision optimization method based on deep learning according to claim 1 is characterized in that: The model architecture design method in S2.1 includes: (1) Encoder part: ①CNN layer construction Construct three sequentially connected CNN layers to process image data; The first layer of CNN, convolution kernel 3×3, number 32, stride 1 and zero padding, input image , after convolution and bias and ReLU activation, the output is , and its calculation formula is: ,in represents the convolution operation, is the bias term; The second layer of CNN uses 5×5 convolution kernels, the number is 64, the stride is 2, no padding, and the input , output ; ; The third layer of CNN uses 7×7 convolution kernel, number 128, stride 2, no padding, input , output , ; ②Pooling layer settings Connect two maximum pooling layers, the pooling kernel is 2×2, the stride is 2, and the first pooling layer input is have to : ; The second pooling layer input have to : ; ③Fully connected layer integration The last fully connected layer converts the output of the pooling layer into a suitable dimension for concatenation with other data. Let the output dimension be , the input is , the weight matrix is , the bias is , then the output ; (2) Data fusion and decoder input preparation: The normalized vehicle status data , the dimension is , and radar data , the dimension is , and the encoder’s fully connected layer output After splicing, the dimension of the vector after splicing is , denoted as ; (3) Decoder part ①Fully connected layer feature fusion The first fully connected layer, weight , bias ,enter , output dimension , activated by ReLU: ; The second fully connected layer has a weight of W d2 , bias b d2 , output dimension : ; ②LSTM layer timing learning Let the first LSTM layer input , the number of hidden units , the relevant weight matrix , , , , bias , , , , the current input is , the hidden state at the previous moment is , the cell state at the last moment is , the Forget Gate , input gate , candidate cell states , cell state update , output gate , hide status updates ; The second LSTM layer is similar, with the number of hidden units set to ; ③Attention mechanism Set the decoder LSTM layer hidden state , encoder output , attention score , the attention distribution is obtained by softmax normalization , the context vector , concatenated with the LSTM hidden state as the output layer input; (4) Output layer: Set weight , bias , the input is the concatenation vector of attention and LSTM output, and the output node corresponds to the number of driving decision categories , the calculation formula is , and the decision probability is obtained by softmax ,in For the The probability of a driving decision.

6. The vehicle intelligent driving decision optimization method based on deep learning according to claim 4 is characterized by: The S2.2 model initialization and hyperparameter setting specifically include: (1) Model initialization: Use Xavier to initialize weights , the formula is , bias Initialized to 0; (2) Hyperparameter setting: The initial value of the learning rate is 0.001, the batch size is 32, the training cycle is 100, the regularization parameter is 0.0001 and L2 regularization is used, and the loss function , =0.0001, is the cross entropy loss.

7. The vehicle intelligent driving decision optimization method based on deep learning according to claim 1 is characterized by: The S3.1 model training process is as follows: (1) Optimization algorithm settings: The model is trained using an optimization algorithm combining stochastic gradient descent and momentum, and the momentum coefficient is set to , let the model parameters be , the learning rate is , the current batch data gradient is , the parameter update formula is ,in is the last parameter update amount, the new parameter , learning rate The initial value is set to 0.001, and the exponential decay learning rate strategy is adopted. The formula is: ,in , every time you pass training cycles, the learning rate decays to the original times; (2) Loss function calculation: Using multi-classification cross entropy loss function , let the model prediction output be , the true label is , the sample size is , whose formula is ,in is the number of driving decision categories. After each training batch of data, the batch loss value is calculated according to this formula; (3) Regular evaluation operations: Every 5 training cycles, the model is evaluated using the validation set data and the accuracy of the model on the validation set is calculated. , recall rate , Value; Accuracy calculation formula ,in Indicates a true example, represents a true counterexample, represents a false positive example, Represents a false negative example; Recall calculation formula ; Value calculation formula .

8. The vehicle intelligent driving decision optimization method based on deep learning according to claim 7 is characterized by: The specific steps of S3.2 model verification and adjustment are as follows: (1) Overfitting and underfitting discrimination: If the model's loss value in the training set continues to decrease, while the performance index of the validation set no longer improves or even decreases, it is considered to be overfitting; if the performance index of the validation set is always low and the loss value decreases slowly, it is considered to be underfitting; (2) Adjustment strategy implementation: When overfitting: Increase the amount of training data and collect more vehicle driving data; Adjust the regularization parameters, such as Regularization parameter Increasing from 0.0001 to 0.001, the loss function becomes ; Use dropout technology to set the dropout probability in the fully connected layer or LSTM layer ; When underfitting: Increase the model complexity, increase the number of neurons or add network layers in the fully connected layer or LSTM layer of the decoder, and pay close attention to the performance of the validation set after adjustment; Adjust the hyperparameters and set the learning rate Increase from 0.001 to 0.01 and increase the training cycle while paying attention to preventing overfitting; (3) Application of visual monitoring: Use the TensorBoard visualization tool to monitor the training process, draw the loss value curve, observe the changes in the loss values ​​of the training set and the validation set with the training cycle, and judge overfitting or underfitting based on this; draw the accuracy curve and parameter distribution curve, the former shows the model learning effect, and the latter reflects the changes in model parameters.

9. The vehicle intelligent driving decision optimization method based on deep learning according to claim 7 is characterized in that: The S4.1 decision optimization strategy is specifically: (1) Model integration and preliminary decision-making The deep learning model is integrated into the intelligent driving system, connected to the vehicle sensor data interface, and receives multi-source data; the model processes the data according to its neural network architecture, outputs the probability values ​​corresponding to multiple driving decisions, and forms a driving decision probability distribution vector ,in is the number of driving decision categories and satisfies ,in Indicates The probability of a driving decision, setting the probability threshold , ,when , corresponding to the decision Become a preliminary candidate decision ; (2) Risk assessment and decision optimization Constructing vehicle surroundings information set ,in Design risk assessment functions for environmental information elements determined based on vehicle dynamics, traffic rules and safe driving criteria , which is based on the environment information set The elements in the formula are constructed, and preliminary candidate decisions are obtained through comprehensive operations on each element. The risk value under the current environment; the risk assessment function has the following form: ,in Based on The weighting factor for determining the importance in the risk assessment, It is for Constructed risk assessment subfunction; based on the risk assessment results, introduce the risk adjustment coefficient , value range , through the formula The probability of each driving decision is adjusted, where After adjustment driving decision probability values, and , thus obtaining optimized driving decisions; (3) Environmental perception feedback mechanism When the vehicle makes a driving decision When the vehicle is in the environment, the environmental perception system monitors the changes in the vehicle's surrounding environment in real time and builds an environmental change information set. ,in For environmental information elements The corresponding change amount; Building a feedback adjustment function This function is constructed based on the inherent logical relationship between the vehicle kinematic model and environmental changes. When the environmental change meets certain conditions or is within a certain range, the formula The probability of each driving decision is adjusted, where It is the adjustment coefficient related to environmental changes, and its value depends on environmental factors. It is based on the feedback adjustment function Function constructed to calculate decision adjustment weights.

Images