CN111667099A - Dynamic target uncertain motion trajectory prediction method based on time granularity improvement - Google Patents

Abstract

The invention belongs to the field of target detection and robot operation, and discloses a dynamic target uncertain motion trajectory prediction method based on time granularity improvement, which comprises the following steps: acquiring original motion trajectory data by using a robot vision processing module, and standardizing the data to be used as a training set; realizing the granularity of the track data through granularity division; selecting a proper weight coefficient of the basis predictor through a time granularity lifting strategy, completing training and constructing a high-precision collaborative prediction model; and (4) obtaining a multi-step high-precision predicted value through rolling iteration and base predictor integration. The method greatly improves the prediction precision and accuracy of the uncertain motion trail of the dynamic target by extracting the effective characteristics of the trail through granularity under the condition of keeping the training time basically unchanged, also enhances the anti-interference capability of the algorithm on noise information, and has important significance for implementing autonomous dynamic operation of a subsequent robot.

Description

Dynamic target uncertain motion trajectory prediction method based on time granularity improvement

Technical Field

The invention belongs to the field of target detection and robot operation, and relates to a dynamic target uncertain motion trajectory prediction method based on time granularity improvement.

Background

The motion trail prediction is an important link for the robot to perform autonomous dynamic operation in the visual environment. The goal of trajectory prediction is to analyze the observed data over a period of time in the past and then provide trajectory information for the next time or a future period of time.

The current research mainly focuses on dynamic operations of known motion models in a structured environment, such as assembly, sorting, stacking and the like of various objects on a production line, and still remains in simple and repetitive mechanical operations. With the rapid development of the fields of big data and artificial intelligence, a fully-automatic intelligent production line based on data analysis and driving is becoming the first choice in the future. However, for such an intelligent production line, the motion trajectories of various objects often have certain incompleteness and uncertainty, and when a robot performs dynamic operation, it is difficult to accurately and effectively acquire the motion trajectories of the objects, so that high-precision trajectory prediction in a certain visual range becomes very important. The current mainstream trajectory prediction method mainly comprises ARIMA and an improved algorithm thereof, Markov prediction and an improved algorithm thereof, and RNN and an improved algorithm thereof. However, the prediction method based on ARIMA is only applicable to standard steady-state timing sequence trajectory data and cannot be applied to an environment with unknown and uncertain values; the Markov-based prediction method has higher computational complexity, so that the consumption of computational resources is higher, and the prediction time is longer; the prediction method based on the RNN, such as the LSTM neural network, has a fast prediction speed, but has poor prediction accuracy and robustness, and is easily affected by noise. These show that a single prediction method often has great limitations, and is difficult to meet the requirements of trajectory prediction precision and robustness of robot dynamic operation in unknown or uncertain environments. Therefore, the invention provides a motion trajectory prediction method based on time granularity improvement, and the core of the motion trajectory prediction method is a multi-LSTM collaborative prediction model based on time granularity. The method has short running time, greatly improves the track prediction precision and provides guarantee for the autonomous operation of a subsequent robot system. It should be noted that time granularity is a general term for a time series data set having a certain length.

Disclosure of Invention

The invention aims to provide a dynamic target uncertain motion track prediction method based on time granularity improvement, which can greatly improve the prediction precision of a single LSTM network through a newly-proposed time granularity improvement mechanism so as to provide accurate high-precision prediction for a dynamic target uncertain motion track.

The invention provides a dynamic target uncertain motion trajectory prediction method based on time granularity improvement, which comprises the following steps of:

s1: acquiring a historical motion track of a target object through a vision processing module of a robot system, and carrying out standardization processing to obtain a track training set G;

s2: carrying out fine-grained division on the trajectory training set, namely enabling G to be { G ═ G₁,G₂,…,G_i,…,G_N}; among them, in the above-mentioned case,

0<i≤N，0<j is less than or equal to H; n represents the number of time granularity sets, and H represents the number of historical tracks;

s3: let the set of basis predictors be { ρ₁,ρ₂,...,ρ_NIs given as { α } corresponding to the set of weight coefficients₁,α₂,...,α_N}; wherein each element in the set of basis predictors represents a standard LSTM network element; respectively inputting the training sets subjected to the fine granularity division in the step S2 into corresponding basis predictors for network training; the training error for each base predictor unit is as follows:

wherein E represents a mathematical expectation; for each base predictor, when the training error is less than 0.5, the training is finished;

s4: according to the training error obtained in the iterative training process in step S3, the weight coefficient corresponding to the basis predictor is calculated using the following formula:

s5: when testing, firstly observing and obtaining the motion trail of a target; if the starting moment of the observation time is s and the finishing moment of the observation is t, the time used for observation is t-s; next, the observed data is subjected to the granularity division in step S2, and is input to the corresponding base predictor, and the predicted value of each base predictor at the next time t +1 is calculated

Is expressed by the formula:

s6: based on the weight coefficient of the base predictor calculated in step S4, the overall predicted value corresponding to the next time t +1 is calculated by the following formula in combination with the predicted value in step S5, and is used as the predicted value corresponding to the next time t +1

S7: after the predicted value of the next moment is obtained, multi-step long-term prediction is carried out through rolling iteration, and the predicted value of the motion track corresponding to the expected moment t + x is obtained:

i.e. predicted value of the track in the future

The method has the advantages that the motion trail is divided into the trail segments with different characteristic attributes by adopting time granularity division, then the segments are respectively subjected to single-base classifier prediction, weight distribution is carried out according to the importance of different trail segments, and finally multi-step period prediction of the motion trail is realized through the cooperative prediction of a multi-base predictor. The method fully extracts the characteristic information of the uncertain motion trail of the dynamic target, greatly improves the prediction precision on the premise of ensuring that the running time meets the requirement, and has a larger development prospect.

Drawings

FIG. 1 is a schematic diagram of trajectory prediction based on time granularity.

FIG. 2 is a flow chart of a dynamic target uncertain motion trajectory prediction system based on time granularity lifting.

FIG. 3 is a comparison of the prediction results of the motion trajectory of a spatial spiral motion object with noise.

Detailed Description

Embodiments of the present invention will be described in further detail with reference to the accompanying drawings 1 to 3

Examples

The invention relates to a dynamic target uncertain motion trajectory prediction method based on time granularity improvement, which comprises the following specific implementation steps of:

s1: acquiring a historical motion track of a target object through a vision processing module of a robot system, and carrying out standardization processing to obtain a track training set G;

s2: carrying out fine-grained division on the trajectory training set, namely enabling G to be { G ═ G₁,G₂,…,G_i,…,G_N}. Among them, in the above-mentioned case,

n denotes the number of time granularity sets and H denotes the number of history tracks.

S3: let the set of basis predictors be { ρ₁,ρ₂,...,ρ_NIs given as { α } corresponding to the set of weight coefficients₁,α₂,...,α_N}. Where each element in the set of base predictors represents a standard LSTM network element. And respectively inputting the training sets subjected to fine granularity division in the S2 into corresponding basis predictors for network training. The training error for each base predictor unit is as follows:

where E represents the mathematical expectation. For each basis predictor, training ends when the training error is less than 0.5.

S4: according to the training error obtained in the iterative training process in the S3, calculating a weight coefficient corresponding to the basis predictor by using the following formula:

s5: when testing, the motion trail of the target needs to be observed and obtained. If the start time of the observation time is s and the end time of the observation is t, the time taken for the observation is t-s. Next, the data obtained by observation is subjected to granularity division as shown in S2, and is input to the corresponding base predictor, and the predicted value of each base predictor at the next time t +1 is calculated

Is formulated as follows. Then, a schematic diagram of the prediction process based on temporal granularity increase is shown in fig. 1.

S6: according to the formula in S4The weight coefficient of the calculated basis predictor is combined with the predicted value in S5, and the overall predicted value corresponding to the next time t +1 can be calculated by using the following formula, and is taken as the predicted value corresponding to the next time t +1 in the method provided by the patent

S7: after the predicted value of the next moment is obtained, multi-step long-term prediction is carried out through rolling iteration, and the predicted value of the motion track corresponding to the expected moment t + x is obtained:

i.e. predicted value of the track in the future

The specific system flow diagram is shown in fig. 2.

S8: the method is also simulated and verified by taking a noisy object with spatial spiral motion as a target, and compared with a standard LSTM algorithm, and the prediction result is shown in FIG. 3. It should be noted that the sampling interval required for observation and prediction is 0.1 s. The core parameters of the standard LSTM network are set as: segmentation window length L60, state vector size Q_stateThe learning rate l is 0.08 at 4. The example method selects the number of basis predictors to be 7, and the corresponding parameters are shown in table 1.

Table 1 detailed parameters of the method of the embodiment (simulation test)

Base predictor	L	Qstate	l	α
					LSTM 1	20	4	0.02	0.08
LSTM 2	40	2	0.008	0.08
					LSTM 3	60	2	0.01	0.12
LSTM 4	80	4	0.006	0.26
					LSTM 5	100	4	0.005	0.14
LSTM 6	120	2	0.004	0.12
					LSTM 7	140	4	0.01	0.20

Fig. 3 shows that, compared with the LSTM algorithm, the method provided by the embodiment is closer to a theoretical trajectory curve, has higher tolerance to noise, and has excellent prediction accuracy and robustness.

S9: selecting 6 different objects, randomly placing the objects on a guide rail running at an uncertain speed, acquiring a motion track required by observation through a vision system of a double-arm cooperative robot system NEXTAGE, predicting the track, further verifying the method, and comparing the method with an ARIMA algorithm and a standard LSTM algorithm. In the embodiment, the observation time is 15s, the prediction time is 20s, and the MAE, the MRE, the RMSE and the accuracy are used as evaluation indexes of the prediction result. For the ARIMA method, the autoregressive coefficient is set to 2, the number of difference terms is 1, and the number of moving average terms is 2. For the standard LSTM method, the parameters are set as: segmentation window length L40, state vector size Q_stateThe learning rate l is 0.04 at 6. Specific parameter settings for the example methods are shown in table 2. Based on the above parameter settings, the comparison ratio of the specific experimental results of each method is shown in table 3.

Table 2 detailed parameters of the method of the example (actual object prediction experiment)

Base predictor	L	Qstate	l	α
					LSTM 1	5	2	0.01	0.15
LSTM 2	10	4	0.005	0.16
					LSTM 3	15	4	0.002	0.41
LSTM 4	20	4	0.001	0.28

TABLE 3 comparison of the predicted results of different objects

As can be seen from Table 3, compared with the traditional ARIMA, LSTM and other methods, the method provided by the invention has higher prediction precision and accuracy, and completely meets the requirements of the dynamic operation of the robot on the track prediction.

Claims (1)

1. The dynamic target uncertain motion trajectory prediction method based on time granularity improvement is characterized by comprising the following steps of:

s1: acquiring a historical motion track of a target object through a vision processing module of a robot system, and carrying out standardization processing to obtain a track training set G;

s2: carrying out fine-grained division on the trajectory training set, namely enabling G to be { G ═ G₁,G₂,…,G_i,…,G_N}; among them, in the above-mentioned case,

0<i≤N，0<j is less than or equal to H; n represents the number of time granularity sets, and H represents the number of historical tracks;

s3: let the set of basis predictors be { ρ₁,ρ₂,...,ρ_NIs given as { α } corresponding to the set of weight coefficients₁,α₂,...,α_N}; wherein each element in the set of basis predictors represents a standard LSTM network element; respectively inputting the training sets subjected to the fine granularity division in the step S2 into corresponding basis predictors for network training; the training error for each base predictor unit is as follows:

wherein E represents a mathematical expectation; for each base predictor, when the training error is less than 0.5, the training is finished;

s4: according to the training error obtained in the iterative training process in step S3, the weight coefficient corresponding to the basis predictor is calculated using the following formula:

s5: when testing, firstly observing and obtaining the motion trail of a target; if the starting moment of the observation time is s and the finishing moment of the observation is t, the time used for observation is t-s; next, the observed data is subjected to the granularity division in step S2, and is input to the corresponding base predictor, and the predicted value of each base predictor at the next time t +1 is calculated

Is expressed by the formula:

s6: based on the weight coefficient of the base predictor calculated in step S4, the overall predicted value corresponding to the next time t +1 is calculated by the following formula in combination with the predicted value in step S5, and is used as the predicted value corresponding to the next time t +1

S7: after the predicted value of the next moment is obtained, multi-step long-term prediction is carried out through rolling iteration, and the predicted value of the motion track corresponding to the expected moment t + x is obtained:

i.e. predicted value of the track in the future

Images