Fact Finder
Top 20
The AKIDA on the BRN website is black. Not sure if it’s the black box you were looking for but I think someone sat on it cause it’s kinda flat.Interestingly, Valeo are at pains to explain that they had developed the algorithms in-house. Well this is a patent they developed for ADAS/AV.
But, and this is a very big butt, they do not claim to have invented the NNs, to the extent that they treat the NNs as "black boxes".
US2021166090A1 DRIVING ASSISTANCE FOR THE LONGITUDINAL AND/OR LATERAL CONTROL OF A MOTOR VEHICLE
VALEO:
Figs 1 to 3 show prior art (known) ADAS arrangements, and the patent then goes on to describe the deficiencies of these systems before describing the invention with reference to Figures 4 to 6.
View attachment 8303
View attachment 8304
[0008] The “online” operation of one known system 3 of this type is shown schematically in FIG. 2. The system 3 comprises a neural network 31 , for example a deep neural network or DNN, and optionally a module 30 for redimensioning the images in order to generate an input image Im′ for the neural network, the dimensions of which are compatible with the network, from an image Im provided by a camera 2 . The neural network forming the image processing device 31 has been trained beforehand and configured so as to generate, at output, a control instruction Scom , for example a (positive or negative) setpoint acceleration or speed for the vehicle when it is desired to exert longitudinal control of the motor vehicle, or a setpoint steering angle of the steering wheel when it is desired to exert lateral control of the vehicle, or even a combination of these two types of instruction if it is desired to exert longitudinal and lateral control.
[0009] In another known implementation of an artificial-intelligence driving assistance system, shown schematically in FIG. 3, the image Im captured by the camera 2 , possibly redimensioned to form an image Im′, is processed in parallel by a plurality of neural networks in a module 310 , each of the networks having been trained for a specific task. Three neural networks have been shown in FIG. 3, each generating an instruction P1 , P2 or P3 for the longitudinal and/or lateral control of the vehicle, from one and the same input image Im′. The instructions are then fused in a digital module 311 so as to deliver a resultant longitudinal and/or lateral control instruction Scom .
[0010] In both cases, the neural networks have been trained based on a large number of image records corresponding to real driving situations of various vehicles involving various humans, and have thus learned to recognize a scene and to generate a control instruction close to human behaviour.
[0011] The benefit of artificial-intelligence systems such as the neural networks described above lies in the fact that these systems will be able to simultaneously apprehend a large number of parameters in a road scene (for example a decrease in brightness, the presence of several obstacles of several kinds, the presence of a car in front of the vehicle and whose rear lights are turned on, curved and/or fading marking lines on the road, etc.) and respond in the same way as a human driver would. However, unlike object detection systems, artificial-intelligence systems do not necessarily classify or detect objects, and therefore do not necessarily estimate information on the distance between the vehicle and a potential hazard.
#####################################################
Figs 4 to 6 relate to zooming in on an item of interest in the field of view and analysing the unzoomed and zoomed images using parallel NNs.
######################################################
[0037] The image processing device 31 a comprises for example a deep neural network.
[0038] The image processing device 31 a is considered here to be a black box, in the sense that the invention proposes to improve the responsiveness of the algorithm that it implements without acting on its internal operation.
[0039] To this end, the invention makes provision to perform, in parallel with the processing performed by the device 31 a, at least one additional processing operation using the same algorithm as the one implemented by the device 31 a, on an additional image formulated from the image Im1 .
[0040] To this end, according to one possible embodiment of the invention, the system 3 comprises a digital image processing module 32 configured so as to provide at least one additional image Im2 at input of an additional image processing device 31 b, identical to the device 31 a and accordingly implementing the same processing algorithm, this additional image Im2 resulting from at least one geometric and/or radiometric transformation performed on the image [#### ie, ZOOM! ####] Im1 initially captured by the camera 2 . In this case too, the system 3 may comprise a redimensioning module 30 b similar to the redimensioning module 30 a, in order to provide an image Im2 ′ compatible with the input of the additional device 31 b.
The algorithm has the effect that the zoomed image can cause the ADAS to apply the brakes earlier than it would in response to the unzoomed image if the leading vehicle applies its brakes.
Well don't aske me - I just live here ...
![Cowboy hat face :cowboy: 🤠](https://cdn.jsdelivr.net/joypixels/assets/6.6/png/unicode/64/1f920.png)
AKIDA BALLISTA