Observing the state of objects from the perspective of an unmanned vehicle - Apollo prediction path revealed_Intefrankly

Observing the state of objects from the perspective of an unmanned vehicle - Apollo prediction path revealed

In the last lesson we have learned how to use sensors to sense the world, the next question is how to predict the object movement state. In this session, we'll learn how Apollo predicts the path of other moving objects based on the state of those objects and the location of the unmanned vehicle.

Forecasting is very important. If you are driving a car in a free space and you find yourself constantly colliding with the cars around you, it may be because you are obstructing the path of other vehicles. thereforeIt is important to predict future environmental conditions around you.

Forecast Profile

Unmanned vehicles will travel between many objects, many of which are themselves moving all the time, like other cars, bicycles and pedestrians. Unmanned vehicles need Predicting these object behaviors is what ensures that the best decisions are made by unmanned vehicles.

The behavior of an object can be predicted by generating a path. In the picture below this car turns to the right and starts to slow down before the ramp, which is one of the paths we predicted for this car.

In the larger context of vehicle travel, we make similar predictions for all other objects on the road, and these predictions come together over time to form Predicting the path.During each time period， We will recalculate for each vehicle merge Predicting their newly generated paths， These predicted paths for unmanned vehicles Decision-making in the planning phase Necessary information was provided.

The predicted paths are topicality harmony accuracy The requirements of.

Real-time means that we want the algorithm to have as short a latency as possible. If a car is going60 kilometre/ hours， So it's every0.25 It will travel in seconds5 meter (classifier)， So we need to make sure that before driving the unmanned vehicle ahead5 No obstacles in meters， merge and can be safely passed。

The next target is accuracy。 If the predictions show that cars in adjacent multiple lanes are trying to merge into our lane, then the thing we need to do is slow down. And in the other case, if it is predicted that cars in adjacent multiple lanes will stay in their lanes, then we need to make predictions that should remain as accurate as possible so that they can help the unmanned car make better decisions.

The prediction module should also be able to learn new behaviors. When there is a lot of traffic on the road, the road conditions become very complex and trying to develop a static model for each scenario at this point is a nearly impossible task. This is why when faced with a problem like this, we need to The prediction module is capable of learning new behaviors. With this approach, we can train using multiple sources of data, which allows the algorithm to improve its predictive power over time.

Common prediction methods

Now we are going to learn the architecture of Apollo platform for prediction. Model-based forecasting together with Data-driven forecasting are the two basic types of predictions.

To make predictions more easily in specific environments (e.g., highways, etc.), we can use this almost infinite number of trajectories to limit the prediction problem. This will make it easier for the car to store the discontinuous lane information it retains at any point in time, so we will simplify the prediction and make it easy to control. moreover Forecasting will also be at the heart of decision making that Because you'll make better decisions if you know what to expect.

Model-based forecasting

Suppose the unmanned car comes to a T-intersection and sees a car coming from the left, at which point it is not clear whether the car is going to turn right or go straight.

Using a model-based approach, we can construct two candidate predictive models for this scenario: one model describes the vehicle performing Right turns, indicated by green tracks. Another model describes the car Continue forward, indicated by the blue track. At this moment, we consider the probability of either model occurring to be the same, so we have two candidate models, each with its own trajectory.

Two candidate prediction models

We will continue to observe the movement of the mobile car， Look at it. together with Which track is a better match： If you see a vehicle start to change lanes to the left， We'll be more confident that the vehicle will eventually go straight， on the other hand， If you see a car holding forward in the right turn lane， We would be more inclined to predict a right turn for that vehicle。 Here it is. The working principle of model-based prediction methods.

Data-driven forecasting

Data-driven forecasting Using machine learning algorithms， Training the model by observing the results。 Once the machine learning model is trained, we can use this model in the real world to make predictions.

Apollo Platform's Architecture for Prediction

The advantages of a data-driven approach are. The more training data there is, the better the model works.

The advantages of the model-based approach are. Intuitive. merge It combines our existing knowledge of physics and the laws of traffic with knowledge of many aspects of human behavior.

Which of the above predictions do you think is better?

Apollo lane based prediction

Apollo offers a new type of software called A lane sequence-based approach that In order to build a sequence of lanes, the road needs to be first divided into sections, each covering a region where vehicle movement can be easily described.

For example, the image above shows an intersection in a partial area. For prediction purposes, we are more interested in how vehicles transition within these zones than in their specific behavior within a particular zone. We can divide the behavior of a vehicle into a finite set of pattern combinations and describe these pattern combinations as Lane Sequence.

For example, the motion of a straight ahead car can be described as a lane sequence of 0-1-3-7.

Lane sequence: 0-1-3-7

Obstacle status

To predict the motion of an object， We also need to know the state of the object。 As a human driver， We need to observe the orientation of an object while driving by、 location、 speed harmony Acceleration to predict what the object will do。

Observation of the orientation, position, velocity and acceleration of objects from a human perspective

This is the same how a driverless car observes the state of an object. In addition to position, velocity, orientation and acceleration, unmanned vehicles need to consider the position of objects within the lane segment. For example, the prediction module considers the longitudinal and lateral distances from the object to the lane line segment boundary.

Orientation of objects from the perspective of an unmanned vehicle、 location、 speed harmony acceleration

The prediction module also contains status information for previous time intervals in order to make more accurate predictions.

This session opens a new chapter in the prediction module of the Unmanned Vehicle Technology course， Our initial understanding harmony Learned what exactly constitutes a forecast， And what are the different kinds of predictions。 Feel free to discuss your thoughts on the "predictions" for unmanned cars in the comments section below!

In the next lesson we will continue to learn about predicting target lanes and more in the prediction module.

Recommended>>
1、Seven Day Crash Course Mini Program Himalaya
2、Google releases AutoMLVision fully automated training AI without writing code
3、Facilitating international technology convergence ECBT participates in the worlds first CephaloconAPAC 2018 Summit
4、Big earthquake Google suddenly announced Huawei is confused Baidu and Drip are going to cry
5、Establishing a scientific risk management framework for commercial banks

已推荐到看一看 和朋友分享想法
最多200字，当前共 发送

已发送

确定
分享你的想法...
取消

确定
最多200字，当前共

发送中