1. Introduction _a forecast
Unmanned shuttle car is traveling among many objects, many of which the object itself is moving, unmanned vehicles need to predict the behavior of these objects, so as to ensure that no car to make the best decisions.
In the Apollo, by generating a path to predict the behavior of the object, such as shown in the following figure, with the ramp in the car as the right and begins to decelerate, then the possible path is shown below:
in motion, will Apollo make all the other objects similar trajectory prediction, which together form the predicted path over a period of time, within each period of time, Apollo will re-calculate the predicted path of their newly generated for each car, these predictions decision unmanned vehicles made by the planning stages provides an important reference.
2. Introduction to predict _b
Predicted path has real-time requirements, delay algorithm that is as low as possible;
Another goal is accuracy, such as the adjacent lane car in the end is to be incorporated into the lane or straight.
Prediction module should also be able to learn new behaviors. When there are a lot of vehicles on the road, the situation will become complicated to develop a model for each static scene is an impossible task, which is why we need when we encounter such a situation, the prediction module able to learn new behaviors. In this way, we can make use of multiple sources of data for training, so you can make the algorithm over time, and to enhance the ability to predict.
3. Sebastian introduce forecast
Prediction can be said is the core of decision-making.
4. Different ways of prediction
Prediction There are two main methods: based on the prediction model and the data-driven prediction .
Model-based prediction algorithm is actually MPC class. Model predictive control (Model Predictive Control) refers to a class of algorithms, the measurement information periodically based on the frame line when solving a finite time open-loop optimization problem, and the front portion of the control sequence applied to the result of the controlled object. Depending on the model, divided into Dynamic Matrix Control (DMC), model algorithmic control (the MAC), Generalized Predictive Control (GPC). In intelligent driving directions, focusing on model predictive control based on the state space model.
Shown below is a typical predicted vehicle model is straight or turn right, intended to determine the vehicle driving operation of the vehicle according to the next.
Data-driven prediction is the use of a machine learning algorithm to train the model by observing the results, once the machine learning model training good, we can use this model in the real world to make predictions.
Based on the advantages of model prediction is that it is intuitive, and a wide range of knowledge combined with our present understanding of physics and traffic laws as well as human behavior.
Data-driven prediction has the advantage of more training data, the better the model results.
5. Based on the predicted lane
Apollo is provided a method for a lane called sequence-based sequence in order to establish the lane, the road is divided into a plurality of first portions, each covering a portion of the area of movement of a vehicle for ease of description, for example, FIG intersection is a partial region.
In order to predict, we are more concerned about is how to convert vehicles in these areas, rather than specific behavior within a certain region. We can conduct a vehicle classified as a combination of a limited set of patterns, and these patterns described as a combination of lane sequence. For example, the movement of the car straight lane can be described as a sequence is 0-1-3-7.
6. obstacle state
In order to predict the motion of the object, we also need to know the state of the object, when we drive vehicle, as human beings, we have to predict what it will do by observing a surface on the object, the position, velocity and acceleration, which is also a no how do people and vehicles to observe the state of the method of the object.
In addition to orientation, position, velocity, acceleration, no need to consider the position of the vehicle within a lane segment of the object. For example, consider the prediction module from the object to the longitudinal and lateral distance of the lane boundary line.
Prediction module further comprises status information of the previous time interval, in order to make more accurate predictions.
7. predicted target lane
We use the lane target sequence framework for object generation track on the road, this is a complicated issue, so we start with a simple question, we have to predict the transition between the lane line, we assume in the lane segment 0 detected a car, and we will predict how it will travel in the next few periods, there are two obvious choices: it can stay in the car segment 0 and then turn to the right; or the car may turn to section 1 then straight. At this point, we will predict the problem in order to select the conversion problem.
We then be selected by probability calculated for each lane sequence. We need a model, and vehicle status lane section as inputs, the model for providing vehicle lane probability of each sequence may be used.
We hope that our model is able to learn new behaviors and should therefore be trained to use the model of empirical observations. In training, we will provide a real vehicle behavior to the model, including not only state-lane sections and objects, which also include sequences vehicle lanes final choice. With the record increase over time, the model can be more independent line iterative, and constantly improve accuracy. Each record will consist of a lane segment of interrelated states observed object and the object sequence followed. At each time point, the object takes up a section and having a specific state. The entire record by the relevant state a series of lane segments and composition of the object.
8. Recurrent Neural Networks
Recurrent Neural Network (recursive neural network) proposed in 1990, is considered the promotion of recurrent neural networks (recurrent neural network), and recurrent neural networks with variable topology and the weight sharing, is used to machine learning task contains structural relationships .
RNN basic structure:
9. recurrent neural network to predict the target lane applications
Apollo RNN use to build a model to predict the target vehicle lane.
Apollo sequence to provide a lane RNN model, provide other related objects RNN model state, Apollo connecting the two outputs are fed into the RNN another neural network, the neural network estimates the probability of each sequence of the lane, having lane sequence predicted the highest probability is that we will follow the sequence of the target vehicle.
To train the network, we use the existing records, each record contains a sequence lane, state of related objects and a label that indicates whether the object is a particular lane to follow this sequence. In training, we compare the network output and the true value of the tag, and use the back-propagation to train the network.
10. The trajectory generating
The final step is to predict the trajectory generation, once we predicted lane sequence of objects, we can predict the trajectory of the object. Between points A and B, there are numerous tracks, then it should be how to predict the most likely trajectory of it?
We can start by setting constraints to remove most of the candidate trajectory, vehicle dynamic constraints are obvious, in practice, it is generally used to polynomial fit.
11. Summary of course
This lesson we describe how to convert the complex movement of the vehicle to the lane changing sequence, which can greatly reduce the number of scenes; then, we describe how to use the existing lane sequence of observations indicate, training a neural network to predict ; Finally, we will lane vehicle dynamics and predicted sequence combined generating trajectory prediction for each object.
