Peer vehicle center line determination system and method

The invention relates to a control system for autonomous driving of a vehicle, in particular to a control system for autonomous driving of a transport vehicle, as e.g. a truck and trailer combination.

The autonomous operation of transport vehicles is a new field of inventions. Highly sophisticated functions require high-end hardware infrastructure including different types of sensors and perception technologies. Until now, SAE Automation level 2 systems require presence and attention of a driver. SAE Automation level 3 systems should manage autonomous driving without continuous attention of the driver. This requires interaction with peer vehicles.

Until now, in particular in the field of commercial vehicles and trucks, no satisfying solution or consideration which is able to handle such interaction situations and which is able to carry out minimal risk safety manoeuvre even in the environment of automated or man driven peer vehicles has been known.

Peer vehicle behaviour prediction and pre-emptive adaption to peer vehicle behaviour requires detailed information on the peer vehicle's position, in particular the peer vehicle's lane relative position. Different attempts have been made to determine the peer vehicle's lane relative position. According to one approach, the position of the peer vehicle was determined by a single shot object detector. A single shot detector is a method for detecting objects in pictures recorded by a forward facing camera of the autonomous vehicle by using a single neural network. Such single shot object detectors are easy to train and easy to integrate into systems like control systems for autonomous driving of a vehicle. However, position determination by single shot object detectors is not sufficiently accurate due to a lack of depth information of camera images and varying viewing angles of the peer vehicle. Document US 2007022 1822 A1 teaches a method of tail light detection using such conventional methods.

Thus, no system and method is known providing a sufficiently accurate determination of a peer vehicle's lane relative position.

Therefore, the object underlying the invention is to enable a more accurate determination of a peer vehicle's position.

The object is achieved by a control system according to claim 1 and a control method according to claim 10. Advantageous further developments are subject-matter of the dependent claims.

The accurate determination of the peer vehicle's position can be based on tail light recognition of that vehicle and by a determination of the peer vehicle's center line based on the recognized tail lights.

Peer vehicles have symmetrical features that can be used to determine the peer vehicle's center line. A common feature representing such symmetrical features are the peer vehicle's tail lights. These can be seen on image sequences of a front facing camera providing a video stream. In autonomous vehicle systems, there are multiple sensors that map the environment of the autonomous vehicle. These sensors include LiDARs, cameras, thermal cameras, ultrasonic sensors or others. Optical cameras provide a video stream of the environment. On this video stream, objects can be detected e.g. using deep neural networks and conventional feature extracting models. Vehicle tail light recognition can be treated as an application of video acquisition and thus is susceptible to spatial image recognition. Such spatial image recognition is i.a. susceptible to deep learning approaches. As an example, it is proposed to provide a split artificial neural network architecture consisting of two parts, a feature extractor subnetwork and two symmetrical tail light detector branches. The feature extractor subnetwork is trained on bounding boxes annotating the tail lights on cropped images of peer vehicles and acts as a semantic segmentator and classifies each pixel in the input image. In this specific case, the semantic segmentator provides a heat map by determining a probability for each pixel in the input image that the given pixel is showing a tail light. The right tail light detector subnetwork and the left tail light detector subnetwork receive this heat map concatenated to the original input image. Using this input information, the right tail lights detector subnetwork and the left tail lights detector subnetwork determines a probability distribution over the width of the image for the inner edge of the right tail light and the left tail light. Under consideration of the highest probability peaks, from these distributions, a center line position of the peer vehicle can be determined by averaging the respective positions of the highest probability peaks of the right tail light and the left tail light.

Disclosed is a control system for autonomous driving of a vehicle, configured to determine the center line of a peer vehicle comprising a sensor that is configured to record a pair of one and another symmetrical features of the peer vehicle that are symmetrical to a virtual center line of the peer vehicle and to provide input information, a symmetrical features detector means (11, 12) configured to extract spatial features of one symmetrical feature (5) and the other symmetrical feature (6) from the input information of the sensor, wherein the center line of the peer vehicle is determined to be the average position between the extracted spatial features of the one symmetrical feature and the extracted spatial features of the other symmetrical feature. The accurate peer vehicle center line detection enables accurate position measurement of the peer vehicle. Projecting e.g. the ground base point of the center line from pixel space of a camera sensor to the autonomous vehicle's coordinate-system gives the exact location of the peer vehicle, while if only a pixel region is available for the detected peer vehicle e.g. as a bounding box or segmentation mask, the peer vehicle's calculated position in the world remains vague.

Preferably, the pair of symmetric features is a left tail light and a right tail light or a left front light and a right front light, wherein the extracted spatial features of the symmetrical features is an inner edge of the left tail light, the right tail light, the left front light or the right front light and the first symmetrical feature detector artificial neural network cell determines an inner edge of the left vehicle light, the second symmetrical feature detector artificial neural network cell determines an inner edge of the right vehicle light. Preferably, the sensor is a camera for visible or non-visible light and the input information is an image of a peer vehicle.

Preferably, the spatial features extraction artificial neural network cell generates a heat map containing probability values for each pixel of the input image and the heat map is concatenated with the input image and used as input for the first symmetrical feature detector artificial neural network cell and the second symmetrical feature detector artificial neural network cell.

Preferably, the first symmetrical feature detector artificial neural network cell is configured to determine a probability distribution over the width of the input image for the spatial features of the one symmetrical feature, and the second symmetrical feature detector artificial neural network cell is configured to determine a probability distribution over the width of the input image for the spatial features of the other symmetrical feature.

Preferably, the at least one spatial features extraction artificial neural network cell is a convolutional neural network cell.

Preferably, the artificial neural network is a deep neural network, in particular an artificial neural network trained by deep learning mechanisms.

Disclosed is further a control method for autonomous driving of a vehicle comprising the steps: acquisition of information on a pair of one and another symmetrical features of a peer vehicle by a sensor, providing the information acquired in the acquisition step as input to an artificial neural network having a split neural network architecture, extracting spatial features from the information by a spatial features extraction artificial neural network cell, extracting spatial features of the one symmetrical feature from the input information in a separate artificial neural network cell, extracting spatial features of the other symmetrical feature from the input information in a separate artificial neural network cell, determining the center line of the peer vehicle by taking the average position between the extracted spatial features of the one symmetrical feature and the extracted spatial features of the other symmetrical feature.

Preferably, the pair of symmetric features is a left tail light and a right tail light or a left front light and a right front light, wherein the extracted spatial features of the symmetrical features is an inner edge of the left tail light, the right tail light, the left front light or the right front light and the first symmetrical feature detection step determines an inner edge of the left vehicle light, the second symmetrical feature detection step determines an inner edge of the right vehicle light.

Preferably, the sensor is a camera for visible or non-visible light and the input information is an image of a peer vehicle.

Preferably, the spatial features extraction step generates a heat map containing probability values for each pixel of the input image and the heat map is concatenated with the input image and used as input for the first spatial features of symmetrical feature extraction step and the second spatial features of the symmetrical feature extraction step.

Preferably, the first spatial features of symmetrical feature extraction step determines a probability distribution over the width of the input image for the spatial features of the one symmetrical feature, and the second spatial features of symmetrical feature extraction step determines a probability distribution over the width of the input image for the spatial features of the other symmetrical feature.

Preferably, the at least one spatial features extraction artificial neural network cell is a convolutional neural network cell.

Preferably, the artificial neural network is a deep neural network, in particular an artificial neural network trained by deep learning mechanisms.

Below, the invention is elucidated by means of embodiments referring to the attached drawings. Fig. 1 exhibits a vehicle including and implementing a first embodiment of the invention.

Fig. 2 exhibits a block diagram of an artificial neural network according to the invention.

Fig. 3 exhibits a flow chart of an embodiment according to the invention.

Fig. 4 exhibits input and output information for the artificial neural network according to Fig. 2.

Fig. 1 exhibits a vehicle 1 including a control system 2 according to the invention. In this embodiment, the vehicle 1 is a tractor of a trailer truck. However, any appropriate vehicle can be equipped with the control system 2.

The control system 2 is provided with an environment sensor architecture comprising a first sensor 3. In the embodiment, first sensor 3 is a visual light video camera directed towards the front of the vehicle 1. However, instead of a video camera, first sensor 3 may be embodied by any sensor that is capable of discerning a right and a left tail light of a leading vehicle like ir-cameras or other optical sensors. In the embodiment, the sensor 3 is a camera attached to the cabin of the vehicle 1. The first sensor 3 records the scene in front of the vehicle land is directed towards a leading peer vehicle 4 having a left tail light 5 and a right tail light 6. A dotted line in Fig. 1 represents a virtual center line 7 of the leading peer vehicle.

The control system 2 comprises an artificial neural network 8 corresponding to the block diagram of Fig. 2. The artificial neural network 8 is of a split neural network architecture and contains several cells. The video stream recorded by the first sensor 3 is fed to a first convolutional neural network cell, specifically a cropped vehicle image 15 determination cell 9 as it can be seen in Fig. 4 of a leading peer vehicle 4 . In the embodiment, a bounding box detector (YOLO - "you only look once" approach or region-based convolutional neural networks) detects peer vehicles from the video stream. Alternatively, a semantic segmentator can be used to detect peer vehicles from the video stream. The detected regions of the video stream containing a peer vehicle are cropped by using image processing techniques known in the art. The cropped vehicle image 15 as output of the first convolutional neural network cell 9 is fed to a second convolutional neural network cell, specifically a feature extractor neural network cell 10 that extracts the spatial features of the cropped vehicle image 15 to determine probabilities of positions of the leading peer vehicle's tail lights, i.e. the left turn light 5, and the right turn light 6. The feature extractor cell 10 is trained on bounding boxes annotating the tail lights on cropped images of peer vehicles and acts as a semantic segmentator and classifies each pixel in the input image. In this specific case, the semantic segmentator provides a heat map 16 by determining probabilities for each pixel in the input cropped vehicle image 15 that the given pixel is showing a tail light.

As it can be seen in Fig. 4. the heat map 16 exhibits a left area 17 and a right area 18 of increased probability of pixels representing a back light The heat map 17 output of the feature extractor neural network cell 10 is concatenated with the cropped vehicle image 15 and fed to a left tail light detector cell 11 and a right tail light detector cell 12. The left tail light detector cell 11 determines a probability distribution 19 over the width of the cropped vehicle image 15 for the inner right edge of the left tail light 5. The right tail light detector cell 12 determines a probability distribution 20 over the width of the cropped vehicle image 15 for the inner left edge of the right tail light 6. The output of the left tail light detector cell 11 and the right tail light detector cell 12 are fed to an average position determination cell 13. The average position determination cell 13 may be implemented as a separate artificial neural network cell or may be implemented as an arithmetic unit and is configured to determine the average position of the left tail light's inner edge and the right tail light's inner edge. This average represents a peer vehicle's center line 21.

Fig. 3 exhibits a flow chart of a method according to the invention. In a recording step S1, a video stream of the leading peer vehicle 4 is recorded by the first sensor 3.

In a cropped vehicle extraction step S2, a first video camera frame of the video stream is fed to the artificial neural network 8. In the first convolutional neural network cell 9, a cropped vehicle image of the leading peer vehicle is determined. In a spatial feature extraction step S3, the cropped vehicle image is fed to the feature extraction neural network cell 10. In the feature extraction neural network cell 10, spatial features, i.e. the location of the back lights are extracted. Acting as the semantic segmentator, a heat map 16 is provided by determining probabilities for each pixel in the input cropped vehicle image 15 that the given pixel is showing a tail light.

In a left tail light detection step S4, the heat map 16 is concatenated with the cropped vehicle image 15 and fed to the left tail light detector cell 11 to determine the inner right edge of the left tail light 5. A probability distribution 19 over the width of the cropped vehicle image 15 for the inner edge of the left tail 5 light is generated.

In parallel to step S4, in a right tail light detection step S5, the heat map 16 is concatenated with the cropped vehicle image 15 and fed to the right tail light detector cell 12 to determine the inner left edge of the right tail light 6. A probability distribution 20 over the width of the cropped vehicle image 15 for the inner edge of the right tail light 6 is generated.

In an average tail light position determination step S6, the average position of the left tail light's inner edge and the right tail light's inner edge is determined and considered as the peer vehicle's center line. I.e. taking the highest probability peak from these distributions and averaging their respective positions, the center line position of the peer vehicle 5 is determined, relative to the detection bounding box.

The invention was described by means of an embodiment. The embodiment is only of explanatory nature and does not restrict the invention as defined by the claims. As recognizable by the skilled person, deviations from the embodiment are possible without deviating from the invention according to the scope of the claimed subject-matter.

In the above embodiment, the control system 2 is based on neural networks. However, any system is applicable that allows detection and discrimination of symmetrical features of a peer vehicle. In the above embodiment, the neural network has a split architecture. However, this serves for improvements of the invention and the center line determination based on recognition of symmetrical features of the peer vehicle also can be achieved by applying a monolithic artificial neural network.

In the above embodiment, the center line determination is based on cropped peer vehicle images. It could also be achieved by algorithms that process an entire camera image frame potentially containing several peer vehicles providing center lines for all peer vehicles in a single shot manner.

In the above embodiment, the first sensor 3 is a front facing sensor observing a lead peer vehicle. In addition or instead, a second sensor 3' may be applied facing to the rear of the vehicle 1 to determine the indicator light state of a trailing peer vehicle. By that, the center line of a trailing peer vehicle may be recognized.

In this document, the term "center line" is meant to describe a hypothetical vertical line that represents the symmetrical center in a left-right direction of a vehicle's rear face or front face. It thus represents the projection of the vehicle's longitudinal vertical symmetry plane to the vehicle's rear face or front face.

In this document, the terms "and", "or" and "either ... or" are used as conjunctions in a meaning similar to the logical conjunctions "AND", "OR" (often also "and/or") or "XOR", respectively. In particular, in contrast to "either ... or", the term "or" also includes occurrence of both operands.

Method steps indicated in the description or the claims only serve an enumerative purpose of the required method steps. They only imply a given sequence or an order where their sequence or order is explicitly expressed or is - obvious for the skilled person - mandatory due to their nature. In particular, the listing of method steps do not imply that this listing is exhaustive. In the claims, the word “comprising” does not exclude other elements or steps, and the indefinite article “a” or “an” does not exclude a plurality and has to be understood as "at least one".

LIST OF REFERENCE SIGNS

1 vehicle

2 control system

3 first sensor

3' second sensor peer vehicle

5 left tail light of peer vehicle right tail light of peer vehicle virtual center line of peer vehicle

8 artificial neural network

9 cropped vehicle image determination convolutional neural network cell

10 feature extractor neural network cell

11 left tail light detector neural network cell

12 right tail light detector neural network cell

15 cropped vehicle image

16 back light pixel probability heat map

17 left area of increased probability for tail light

18 right area of increased probability for tail light

19 left tail light probability distribution 0 right tail light probability distribution 1 binary back light pixel probability heat map. 1 recording step 2 cropped vehicle extraction step 3 spatial feature extraction step 4 left tail light detection step 5 right tail light detection step 6 average tail light position determination step