EP4690825A1

EP4690825A1 - Sensor device and method for operating a sensor device

Info

Publication number: EP4690825A1
Application number: EP24708834.7A
Authority: EP
Inventors: Dimche KOSTADINOV; Ryoji Ikegaya; Andreas AUMILLER; Christian Peter BRÄNDLI
Original assignee: Sony Advanced Visual Sensing AG; Sony Semiconductor Solutions Corp
Current assignee: Sony Advanced Visual Sensing AG; Sony Semiconductor Solutions Corp
Priority date: 2023-03-27
Filing date: 2024-03-07
Publication date: 2026-02-11
Also published as: WO2024199933A1

Abstract

A sensor device (10) comprises a vision sensor (1010) that comprises a pixel array (1011) having a plurality of pixels (51) each being configured to receive light and to perform photoelectric conversion to generate an electrical signal, based on which electrical signals a data stream is formed and output by the vision sensor (1010), a modulation unit (1020) that is configured to receive portions of the data stream and to modulate said portions of the data stream to generate modulated data portions, and a processing unit (1030) that is configured to receive the modulated data portions and to carry out a predetermined processing of the modulated data portions. Here, the processing unit (1030) is configured to use a first machine learning module (1035) to carry out the predetermined processing which first machine learning module (1035) has been trained such as to optimize the predetermined processing together with a second machine learning module (1025) used by the modulation unit (1020) to generate the modulated data portions.

Description

SENSOR DEVICE AND METHOD FOR OPERATING A SENSOR DEVICE

FIELD OF THE INVENTION

The present technology relates to a sensor device and a method for operating a sensor device, in particular, to a sensor device and a method for operating a sensor device that allows an improved processing of sensor data.

BACKGROUND

Presently, sensor data obtained in imaging systems like active pixel sensors, APS, and dynamic/event vision sensors, DVS/EVS, are further processed to give estimates on the observed scenes. This is often done by using machine learning algorithms that are trained to fulfill certain tasks in order to generate the desired estimates. Here, it is sometimes difficult to extract meaningful and useful information from the sensor data to fulfill a given task in an efficient manner regarding processing resources and energy consumption.

Improved sensor devices and methods for operating these sensor devices are desirable that mitigate this problem.

SUMMARY OF INVENTION

To this end, a sensor device is provided that comprises a vision sensor that comprises a pixel array having a plurality of pixels each being configured to receive light and to perform photoelectric conversion to generate an electrical signal, based on which electrical signals a data stream is formed and output by the vision sensor, a modulation unit that is configured to receive portions of the data stream and to modulate said portions of the data stream to generate modulated data portions, and a processing unit that is configured to receive the modulated data portions and to carry out a predetermined processing of the modulated data portions; Here, the processing unit is configured to use a first machine learning module to carry out the predetermined processing which first machine learning module has been trained such as to optimize the predetermined processing together with a second machine learning module used by the modulation unit to generate the modulated data portions.

Further, a method for operating an according sensor device is provided, which comprises: by a vision sensor that comprises a pixel array having a plurality of pixels each being configured to receive light and to perform photoelectric conversion to generate an electrical signal, forming and outputting a data stream based on the generated electrical signals; by a modulation unit, receiving portions of the data stream and modulating said portions of the data stream to generate modulated data portions by using a second machine learning module; by a processing unit, receiving the modulated data portions and carrying out a predetermined processing of the modulated data portions by using a first machine learning module; and training the first machine learning module such as to optimize the predetermined processing together with the second machine learning module. By modulating the sensor data by a machine learning module that is trained together with the machine learning module that carries out the predetermined processing, i.e. the desired task, it can be ensured that data input for the predetermined processing is optimized for the task at hand. In particular, modulating the sensor data helps to bring the data in an optimized format. Further, by the modulation it is in principle possible to supplement the sensor data with additional content in a “hidden” manner, i.e. a manner not perceivable by an observer of the resulting image. Training the two involved machine learning modules together provides an effective manner to achieve a well working embedding of hidden content in the sensor data, while at the same time optimize it for a predetermined task.

BRIEF DESCRIPTION OF DRAWINGS

Fig. 1 is a schematic diagram of a sensor device.

Fig. 2 is a schematic block diagram of a sensor section.

Fig. 3 is a schematic block diagram of a pixel array section.

Fig. 4 is a schematic circuit diagram of a pixel block.

Fig. 5 is a schematic block diagram illustrating of an event detecting section.

Fig. 6 is a schematic circuit diagram of a current-voltage converting section.

Fig. 7 is a schematic circuit diagram of a subtraction section and a quantization section.

Fig. 8 is a schematic diagram of a frame data generation method based on event data.

Fig. 9 is a schematic block diagram of another quantization section.

Fig. 10 is a schematic diagram of another event detecting section.

Fig. 11 is a schematic block diagram of another pixel array section.

Fig. 12 is a schematic circuit diagram of another pixel block.

Fig. 13 is a schematic block diagram of a scan-type sensor device.

Fig. 14 is a schematic block diagram of a sensor device.

Figs. 15A and 15B are modulation examples . Fig. 16 is a schematic illustration of a process flow of a method for operating a sensor device.

Fig. 17 is a schematic illustration of a processing of event data.

Fig. 18 is a schematic block diagram of a vehicle control system.

Fig. 19 is a diagram of assistance in explaining an example of installation positions of an outside-vehicle information detecting section and an imaging section.

Fig. 20A and 20B are schematic illustrations of a mobile device and a head mounted display comprising a sensor device.

DETAILED DESCRIPTION

The present disclosure is directed to mitigating problems related to processing of data of imaging sensors. The solutions to these problems discussed below are applicable to all according sensor types. They are particularly relevant for event based/dynamic vision sensors, EVS/DVS, since the sparsity of the sensor data generated for these sensors allows particular improvements of the efficiency of processing these data. In order to simplify the description and also in order to cover an important application example, the present description is focused therefore without prejudice on EVS/DVS. However, it has to be understood that although in the following reference will be made to the circuitry of EVS/DVS, the discussed solutions can be applied in principle to all pixel-based sensor devices. The discussed sensor devices may be implemented in any imaging sensor setup such as e.g. smartphone cameras, scientific devices, automotive video sensors or the like.

First, a possible implementation of a EVS/DVS will be described. This is of course purely exemplary. It is to be understood that EVSs/DVSs could also be implemented differently.

Fig. 1 is a diagram illustrating a configuration example of a sensor device 10, which is in the example of Fig. 1 constituted by a sensor chip.

The sensor device 10 is a single-chip semiconductor chip and includes a sensor die (substrate) 11, which serves as a plurality of dies (substrates), and a logic die 12 that are stacked. Note that, the sensor device 10 can also include only a single die or three or more stacked dies.

In the sensor device 10 of Fig. 1, the sensor die 11 includes (a circuit serving as) a sensor section 21, and the logic die 12 includes a logic section 22. Note that, the sensor section 21 can be partly formed on the logic die 12. Further, the logic section 22 can be partly formed on the sensor die 11.

The sensor section 21 includes pixels configured to perform photoelectric conversion on incident light to generate electrical signals, and generates event data indicating the occurrence of events that are changes in the electrical signal of the pixels. The sensor section 21 supplies the event data to the logic section 22. That is, the sensor section 21 performs imaging of performing, in the pixels, photoelectric conversion on incident light to generate electrical signals, similarly to a synchronous image sensor, for example. The sensor section 21, however, generates event data indicating the occurrence of events that are changes in the electrical signal of the pixels instead of generating image data in a frame format (frame data). The sensor section 21 outputs, to the logic section 22, the event data obtained by the imaging.

Here, the synchronous image sensor is an image sensor configured to perform imaging in synchronization with a vertical synchronization signal and output frame data that is image data in a frame format. The sensor section 21 can be regarded as asynchronous (an asynchronous image sensor) in contrast to the synchronous image sensor, since the sensor section 21 does not operate in synchronization with a vertical synchronization signal when outputting event data. In particular, the sensor section 21 can output event data with a temporal precision of 10'⁶ s.

Note that, the sensor section 21 may generate and output, other than event data, frame data, similarly to the synchronous image sensor. In addition, the sensor section 21 can output, together with event data, electrical signals of pixels in which events have occurred, as pixel signals that are pixel values of the pixels in frame data.

The logic section 22 controls the sensor section 21 as needed. Further, the logic section 22 performs various types of data processing, such as data processing of generating frame data on the basis of event data from the sensor section 21 and image processing on frame data from the sensor section 21 or frame data generated on the basis of the event data from the sensor section 21, and outputs data processing results obtained by performing the various types of data processing on the event data and the frame data. The logic section 22 may implement the functions of a control unit as described below.

Fig. 2 is a block diagram illustrating a configuration example of the sensor section 21 of Fig. 1.

The sensor section 21 includes a pixel array section 31, a driving section 32, an arbiter 33, an AD (Analog to Digital) conversion section 34, and an output section 35.

The pixel array section 31 includes a plurality of pixels 51 (Fig. 3) arrayed in a two-dimensional lattice pattern. The pixel array section 31 detects, in a case where a change larger than a predetermined threshold (including a change equal to or larger than the threshold as needed) has occurred in (a voltage corresponding to) a photocurrent that is an electrical signal generated by photoelectric conversion in the pixel 51, the change in the photocurrent as an event. In a case of detecting an event, the pixel array section 31 outputs, to the arbiter 33, a request for requesting the output of event data indicating the occurrence of the event. Then, in a case of receiving a response indicating event data output permission from the arbiter 33, the pixel array section 31 outputs the event data to the driving section 32 and the output section 35. In addition, the pixel array section 31 may output an electrical signal of the pixel 51 in which the event has been detected to the AD conversion section 34.

The driving section 32 supplies control signals to the pixel array section 31 to drive the pixel array section 31. For example, the driving section 32 drives the pixel 51 regarding which the pixel array section 31 has output event data, so that the pixel 51 in question supplies (outputs) a pixel signal to the AD conversion section 34.

The arbiter 33 arbitrates the requests for requesting the output of event data from the pixel array section 31, and returns responses indicating event data output permission or prohibition to the pixel array section 31.

The AD conversion section 34 includes, for example, a single-slope ADC (AD converter) (not illustrated) in each column of pixel blocks 41 (Fig. 3) described later, for example. The AD conversion section 34 performs, with the ADC in each column, AD conversion on pixel signals of the pixels 51 of the pixel blocks 41 in the column, and supplies the resultant to the output section 35. Note that, the AD conversion section 34 can perform CDS (Correlated Double Sampling) together with pixel signal AD conversion.

The output section 35 performs necessary processing on the pixel signals from the AD conversion section 34 and the event data from the pixel array section 31 and supplies the resultant to the logic section 22 (Fig. 1).

Here, a change in the photocurrent generated in the pixel 51 can be recognized as a change in the amount of light entering the pixel 51, so that it can also be said that an event is a change in light amount (a change in light amount larger than the threshold) in the pixel 51.

Event data indicating the occurrence of an event at least includes location information (coordinates or the like) indicating the location of a pixel block in which a change in light amount, which is the event, has occurred. Besides, the event data can also include the polarity (positive or negative) of the change in light amount.

With regard to the series of event data that is output from the pixel array section 31 at timings at which events have occurred, it can be said that, as long as the event data interval is the same as the event occurrence interval, the event data implicitly includes time point information indicating (relative) time points at which the events have occurred. However, for example, when the event data is stored in a memory and the event data interval is no longer the same as the event occurrence interval, the time point information implicitly included in the event data is lost. Thus, the output section 35 includes, in event data, time point information indicating (relative) time points at which events have occurred, such as timestamps, before the event data interval is changed from the event occurrence interval. The processing of including time point information in event data can be performed in any block other than the output section 35 as long as the processing is performed before time point information implicitly included in event data is lost.

Fig. 3 is a block diagram illustrating a configuration example of the pixel array section 31 of Fig. 2.

The pixel array section 31 includes the plurality of pixel blocks 41. The pixel block 41 includes the IxJ pixels 51 that are one or more pixels arrayed in I rows and J columns (I and J are integers), an event detecting section 52, and a pixel signal generating section 53. The one or more pixels 51 in the pixel block 41 share the event detecting section 52 and the pixel signal generating section 53. Further, in each column of the pixel blocks 41, a VSL (Vertical Signal Line) for connecting the pixel blocks 41 to the ADC of the AD conversion section 34 is wired.

The pixel 51 receives light incident from an object and performs photoelectric conversion to generate a photocurrent serving as an electrical signal. The pixel 51 supplies the photocurrent to the event detecting section 52 under the control of the driving section 32.

The event detecting section 52 detects, as an event, a change larger than the predetermined threshold in photocurrent from each of the pixels 51, under the control of the driving section 32. In a case of detecting an event, the event detecting section 52 supplies, to the arbiter 33 (Fig. 2), a request for requesting the output of event data indicating the occurrence of the event. Then, when receiving a response indicating event data output permission to the request from the arbiter 33, the event detecting section 52 outputs the event data to the driving section 32 and the output section 35.

The pixel signal generating section 53 generates, in the case where the event detecting section 52 has detected an event, a voltage corresponding to a photocurrent from the pixel 51 as a pixel signal, and supplies the voltage to the AD conversion section 34 through the VSL, under the control of the driving section 32.

Here, detecting a change larger than the predetermined threshold in photocurrent as an event can also be recognized as detecting, as an event, absence of change larger than the predetermined threshold in photocurrent. The pixel signal generating section 53 can generate a pixel signal in the case where absence of change larger than the predetermined threshold in photocurrent has been detected as an event as well as in the case where a change larger than the predetermined threshold in photocurrent has been detected as an event.

Fig. 4 is a circuit diagram illustrating a configuration example of the pixel block 41.

The pixel block 41 includes, as described with reference to Fig. 3, the pixels 51, the event detecting section 52, and the pixel signal generating section 53.

The pixel 51 includes a photoelectric conversion element 61 and transfer transistors 62 and 63.

The photoelectric conversion element 61 includes, for example, a PD (Photodiode). The photoelectric conversion element 61 receives incident light and performs photoelectric conversion to generate charges.

The transfer transistor 62 includes, for example, an N (Negative)-type MOS (Metal-Oxide- Semiconductor) FET (Field Effect Transistor). The transfer transistor 62 of the n-th pixel 51 of the IxJ pixels 51 in the pixel block 41 is turned on or off in response to a control signal OFGn supplied from the driving section 32 (Fig. 2). When the transfer transistor 62 is turned on, charges generated in the photoelectric conversion element 61 are transferred (supplied) to the event detecting section 52, as a photocurrent.

The transfer transistor 63 includes, for example, an N-type MOSFET. The transfer transistor 63 of the n- th pixel 51 of the IxJ pixels 51 in the pixel block 41 is turned on or off in response to a control signal TRGn supplied from the driving section 32. When the transfer transistor 63 is turned on, charges generated in the photoelectric conversion element 61 are transferred to an FD 74 of the pixel signal generating section 53.

The IxJ pixels 51 in the pixel block 41 are connected to the event detecting section 52 of the pixel block 41 through nodes 60. Thus, photocurrents generated in (the photoelectric conversion elements 61 of) the pixels 51 are supplied to the event detecting section 52 through the nodes 60. As a result, the event detecting section 52 receives the sum of photocurrents from all the pixels 51 in the pixel block 41. Thus, the event detecting section 52 detects, as an event, a change in sum of photocurrents supplied from the IxJ pixels 51 in the pixel block 41.

The pixel signal generating section 53 includes a reset transistor 71, an amplification transistor 72, a selection transistor 73, and the FD (Floating Diffusion) 74.

The reset transistor 71, the amplification transistor 72, and the selection transistor 73 include, for example, N-type MOSFETs.

The reset transistor 71 is turned on or off in response to a control signal RST supplied from the driving section 32 (Fig. 2). When the reset transistor 71 is turned on, the FD 74 is connected to a power supply VDD, and charges accumulated in the FD 74 are thus discharged to the power supply VDD. With this, the FD 74 is reset.

The amplification transistor 72 has a gate connected to the FD 74, a drain connected to the power supply VDD, and a source connected to the VSL through the selection transistor 73. The amplification transistor 72 is a source follower and outputs a voltage (electrical signal) corresponding to the voltage of the FD 74 supplied to the gate to the VSL through the selection transistor 73.

The selection transistor 73 is turned on or off in response to a control signal SEL supplied from the driving section 32. When the selection transistor 73 is turned on, a voltage corresponding to the voltage of the FD 74 from the amplification transistor 72 is output to the VSL.

The FD 74 accumulates charges transferred from the photoelectric conversion elements 61 of the pixels 51 through the transfer transistors 63, and converts the charges to voltages.

With regard to the pixels 51 and the pixel signal generating section 53, which are configured as described above, the driving section 32 turns on the transfer transistors 62 with control signals OFGn, so that the transfer transistors 62 supply, to the event detecting section 52, photocurrents based on charges generated in the photoelectric conversion elements 61 of the pixels 51. With this, the event detecting section 52 receives a current that is the sum of the photocurrents from all the pixels 51 in the pixel block 41, which might also be only a single pixel.

When the event detecting section 52 detects, as an event, a change in photocurrent (sum of photocurrents) in the pixel block 41, the driving section 32 turns off the transfer transistors 62 of all the pixels 51 in the pixel block 41, to thereby stop the supply of the photocurrents to the event detecting section 52. Then, the driving section 32 sequentially turns on, with the control signals TRGn, the transfer transistors 63 of the pixels 51 in the pixel block 41 in which the event has been detected, so that the transfer transistors 63 transfers charges generated in the photoelectric conversion elements 61 to the FD 74. The FD 74 accumulates the charges transferred from (the photoelectric conversion elements 61 of) the pixels 51. Voltages corresponding to the charges accumulated in the FD 74 are output to the VSL, as pixel signals of the pixels 51, through the amplification transistor 72 and the selection transistor 73.

As described above, in the sensor section 21 (Fig. 2), only pixel signals of the pixels 51 in the pixel block 41 in which an event has been detected are sequentially output to the VSL. The pixel signals output to the VSL are supplied to the AD conversion section 34 to be subjected to AD conversion.

Here, in the pixels 51 in the pixel block 41, the transfer transistors 63 can be turned on not sequentially but simultaneously. In this case, the sum of pixel signals of all the pixels 51 in the pixel block 41 can be output.

In the pixel array section 31 of Fig. 3, the pixel block 41 includes one or more pixels 51, and the one or more pixels 51 share the event detecting section 52 and the pixel signal generating section 53. Thus, in the case where the pixel block 41 includes a plurality of pixels 51, the numbers of the event detecting sections 52 and the pixel signal generating sections 53 can be reduced as compared to a case where the event detecting section 52 and the pixel signal generating section 53 are provided for each of the pixels 51, with the result that the scale of the pixel array section 31 can be reduced.

Note that, in the case where the pixel block 41 includes a plurality of pixels 51, the event detecting section 52 can be provided for each of the pixels 51. In the case where the plurality of pixels 51 in the pixel block 41 share the event detecting section 52, events are detected in units of the pixel blocks 41. In the case where the event detecting section 52 is provided for each of the pixels 51, however, events can be detected in units of the pixels 51.

Yet, even in the case where the plurality of pixels 51 in the pixel block 41 share the single event detecting section 52, events can be detected in units of the pixels 51 when the transfer transistors 62 of the plurality of pixels 51 are temporarily turned on in a time-division manner.

Further, in a case where there is no need to output pixel signals, the pixel block 41 can be formed without the pixel signal generating section 53. In the case where the pixel block 41 is formed without the pixel signal generating section 53, the sensor section 21 can be formed without the AD conversion section 34 and the transfer transistors 63. In this case, the scale of the sensor section 21 can be reduced. The sensor will then output the address of the pixel (block) in which the event occurred, if necessary with a time stamp.

Fig. 5 is a block diagram illustrating a configuration example of the event detecting section 52 of Fig. 3.

The event detecting section 52 includes a current-voltage converting section 81, a buffer 82, a subtraction section 83, a quantization section 84, and a transfer section 85.

The current-voltage converting section 81 converts (a sum of) photocurrents from the pixels 51 to voltages corresponding to the logarithms of the photocurrents (hereinafter also referred to as a "photovoltage") and supplies the voltages to the buffer 82.

The buffer 82 buffers photovoltages from the current-voltage converting section 81 and supplies the resultant to the subtraction section 83.

The subtraction section 83 calculates, at a timing instructed by a row driving signal that is a control signal from the driving section 32, a difference between the current photovoltage and a photovoltage at a timing slightly shifted from the current time, and supplies a difference signal corresponding to the difference to the quantization section 84.

The quantization section 84 quantizes difference signals from the subtraction section 83 to digital signals and supplies the quantized values of the difference signals to the transfer section 85 as event data.

The transfer section 85 transfers (outputs), on the basis of event data from the quantization section 84, the event data to the output section 35. That is, the transfer section 85 supplies a request for requesting the output of the event data to the arbiter 33. Then, when receiving a response indicating event data output permission to the request from the arbiter 33, the transfer section 85 outputs the event data to the output section 35.

Fig. 6 is a circuit diagram illustrating a configuration example of the current-voltage converting section 81 of Fig. 5.

The current-voltage converting section 81 includes transistors 91 to 93. As the transistors 91 and 93, for example, N-type MOSFETs can be employed. As the transistor 92, for example, a P-type MOSFET can be employed.

The transistor 91 has a source connected to the gate of the transistor 93, and a photocurrent is supplied from the pixel 51 to the connecting point between the source of the transistor 91 and the gate of the transistor 93. The transistor 91 has a drain connected to the power supply VDD and a gate connected to the drain of the transistor 93. The transistor 92 has a source connected to the power supply VDD and a drain connected to the connecting point between the gate of the transistor 91 and the drain of the transistor 93. A predetermined bias voltage Vbias is applied to the gate of the transistor 92. With the bias voltage Vbias, the transistor 92 is turned on or off, and the operation of the current-voltage converting section 81 is turned on or off depending on whether the transistor 92 is turned on or off.

The source of the transistor 93 is grounded.

In the current-voltage converting section 81, the transistor 91 has the drain connected on the power supply VDD side. The source of the transistor 91 is connected to the pixels 51 (Fig. 4), so that photocurrents based on charges generated in the photoelectric conversion elements 61 of the pixels 51 flow through the transistor 91 (from the drain to the source). The transistor 91 operates in a subthreshold region, and at the gate of the transistor 91, photovoltages corresponding to the logarithms of the photocurrents flowing through the transistor 91 are generated. As described above, in the current-voltage converting section 81, the transistor 91 converts photocurrents from the pixels 51 to photovoltages corresponding to the logarithms of the photocurrents.

In the current-voltage converting section 81, the transistor 91 has the gate connected to the connecting point between the drain of the transistor 92 and the drain of the transistor 93, and the photovoltages are output from the connecting point in question.

Fig. 7 is a circuit diagram illustrating configuration examples of the subtraction section 83 and the quantization section 84 of Fig. 5.

The subtraction section 83 includes a capacitor 101, an operational amplifier 102, a capacitor 103, and a switch 104. The quantization section 84 includes a comparator 111.

The capacitor 101 has one end connected to the output terminal of the buffer 82 (Fig. 5) and the other end connected to the input terminal (inverting input terminal) of the operational amplifier 102. Thus, photovoltages are input to the input terminal of the operational amplifier 102 through the capacitor 101.

The operational amplifier 102 has an output terminal connected to the non-inverting input terminal (+) of the comparator 111.

The capacitor 103 has one end connected to the input terminal of the operational amplifier 102 and the other end connected to the output terminal of the operational amplifier 102.

The switch 104 is connected to the capacitor 103 to switch the connections between the ends of the capacitor 103. The switch 104 is turned on or off in response to a row driving signal that is a control signal from the driving section 32, to thereby switch the connections between the ends of the capacitor A photovoltage on the buffer 82 (Fig. 5) side of the capacitor 101 when the switch 104 is on is denoted by Vinit, and the capacitance (electrostatic capacitance) of the capacitor 101 is denoted by Cl. The input terminal of the operational amplifier 102 serves as a virtual ground terminal, and a charge Qinit that is accumulated in the capacitor 101 in the case where the switch 104 is on is expressed by Expression (1).

Qinit = Cl x Vinit (1)

Further, in the case where the switch 104 is on, the connection between the ends of the capacitor 103 is cut (short-circuited), so that no charge is accumulated in the capacitor 103.

When a photovoltage on the buffer 82 (Fig. 5) side of the capacitor 101 in the case where the switch 104 has thereafter been turned off is denoted by Vafter, a charge Qafter that is accumulated in the capacitor 101 in the case where the switch 104 is off is expressed by Expression (2).

Qafter = Cl x Vafter (2)

When the capacitance of the capacitor 103 is denoted by C2 and the output voltage of the operational amplifier 102 is denoted by Vout, a charge Q2 that is accumulated in the capacitor 103 is expressed by Expression (3).

Q2 = -C2 x Vout (3)

Since the total amount of charges in the capacitors 101 and 103 does not change before and after the switch 104 is turned off, Expression (4) is established.

Qinit = Qafter + Q2 (4)

When Expression (1) to Expression (3) are substituted for Expression (4), Expression (5) is obtained.

Vout = -(C1/C2) x (Vafter - Vinit) (5)

With Expression (5), the subtraction section 83 subtracts the photovoltage Vinit from the photovoltage Vafter, that is, calculates the difference signal (Vout) corresponding to a difference Vafter - Vinit between the photovoltages Vafter and Vinit. With Expression (5), the subtraction gain of the subtraction section 83 is C1/C2. Since the maximum gain is normally desired, Cl is preferably set to a large value and C2 is preferably set to a small value. Meanwhile, when C2 is too small, kTC noise increases, resulting in a risk of deteriorated noise characteristics. Thus, the capacitance C2 can only be reduced in a range that achieves acceptable noise. Further, since the pixel blocks 41 each have installed therein the event detecting section 52 including the subtraction section 83, the capacitances Cl and C2 have space constraints. In consideration of these matters, the values of the capacitances Cl and C2 are determined. The comparator 111 compares a difference signal from the subtraction section 83 with a predetermined threshold (voltage) Vth (>0) applied to the inverting input terminal (-), thereby quantizing the difference signal. The comparator 111 outputs the quantized value obtained by the quantization to the transfer section 85 as event data.

For example, in a case where a difference signal is larger than the threshold Vth, the comparator 111 outputs an H (High) level indicating 1, as event data indicating the occurrence of an event. In a case where a difference signal is not larger than the threshold Vth, the comparator 111 outputs an L (Low) level indicating 0, as event data indicating that no event has occurred.

The transfer section 85 supplies a request to the arbiter 33 in a case where it is confirmed on the basis of event data from the quantization section 84 that a change in light amount that is an event has occurred, that is, in the case where the difference signal (Vout) is larger than the threshold Vth. When receiving a response indicating event data output permission, the transfer section 85 outputs the event data indicating the occurrence of the event (for example, H level) to the output section 35.

The output section 35 includes, in event data from the transfer section 85, location/address information regarding (the pixel block 41 including) the pixel 51 in which an event indicated by the event data has occurred and time point information indicating a time point at which the event has occurred, and further, as needed, the polarity of a change in light amount that is the event, i.e. whether the intensity did increase or decrease. The output section 35 outputs the event data.

As the data format of event data including location information regarding the pixel 51 in which an event has occurred, time point information indicating a time point at which the event has occurred, and the polarity of a change in light amount that is the event, for example, the data format called "AER (Address Event Representation)" can be employed.

Note that, a gain A of the entire event detecting section 52 is expressed by the following expression where the gain of the current-voltage converting section 81 is denoted by CGi_og and the gain of the buffer 82 is 1.

A = CGio_gC 1/C2 (Zi_Photo_n) (6)

Here, i_Photo_n denotes a photocurrent of the n-th pixel 51 of the IxJ pixels 51 in the pixel block 41. In Expression (6), E denotes the summation of n that takes integers ranging from 1 to HJ.

Note that, the pixel 51 can receive any light as incident light with an optical fdter through which predetermined light passes, such as a color fdter. For example, in a case where the pixel 51 receives visible light as incident light, event data indicates the occurrence of changes in pixel value in images including visible objects. Further, for example, in a case where the pixel 51 receives, as incident light, infrared light, millimeter waves, or the like for ranging, event data indicates the occurrence of changes in distances to objects. In addition, for example, in a case where the pixel 51 receives infrared light for temperature measurement, as incident light, event data indicates the occurrence of changes in temperature of objects. In the present embodiment, the pixel 51 is assumed to receive visible light as incident light.

Fig. 8 is a diagram illustrating an example of a frame data generation method based on event data.

The logic section 22 sets a frame interval and a frame width on the basis of an externally input command, for example. Here, the frame interval represents the interval of frames of frame data that is generated on the basis of event data. The frame width represents the time width of event data that is used for generating frame data on a single frame. A frame interval and a frame width that are set by the logic section 22 are also referred to as a "set frame interval" and a "set frame width," respectively.

The logic section 22 generates, on the basis of the set frame interval, the set frame width, and event data from the sensor section 21, frame data that is image data in a frame format, to thereby convert the event data to the frame data.

That is, the logic section 22 generates, in each set frame interval, frame data on the basis of event data in the set frame width from the beginning of the set frame interval.

Here, it is assumed that event data includes time point information ti indicating a time point at which an event has occurred (hereinafter also referred to as an "event time point") and coordinates (x, y) serving as location information regarding (the pixel block 41 including) the pixel 51 in which the event has occurred (hereinafter also referred to as an "event location").

In Fig. 8, in a three-dimensional space (time and space) with the x axis, the y axis, and the time axis t, points representing event data are plotted on the basis of the event time point t and the event location (coordinates) (x, y) included in the event data.

That is, when a location (x, y, t) on the three-dimensional space indicated by the event time point t and the event location (x, y) included in event data is regarded as the space-time location of an event, in Fig. 8, the points representing the event data are plotted on the space-time locations (x, y, t) of the events.

The logic section 22 starts to generate frame data on the basis of event data by using, as a generation start time point at which frame data generation starts, a predetermined time point, for example, a time point at which frame data generation is externally instructed or a time point at which the sensor device 10 is powered on.

Here, cuboids each having the set frame width in the direction of the time axis t in the set frame intervals, which appear from the generation start time point, are referred to as a "frame volume." The size of the frame volume in the x-axis direction or the y-axis direction is equal to the number of the pixel blocks 41 or the pixels 51 in the x-axis direction or the y-axis direction, for example.

The logic section 22 generates, in each set frame interval, frame data on a single frame on the basis of event data in the frame volume having the set frame width from the beginning of the set frame interval. Frame data can be generated by, for example, setting white to a pixel (pixel value) in a frame at the event location (x, y) included in event data and setting a predetermined color such as gray to pixels at other locations in the frame.

Besides, in a case where event data includes the polarity of a change in light amount that is an event, frame data can be generated in consideration of the polarity included in the event data. For example, white can be set to pixels in the case a positive polarity, while black can be set to pixels in the case of a negative polarity.

In addition, in the case where pixel signals of the pixels 51 are also output when event data is output as described with reference to Fig. 3 and Fig. 4, frame data can be generated on the basis of the event data by using the pixel signals of the pixels 51. That is, frame data can be generated by setting, in a frame, a pixel at the event location (x, y) (in a block corresponding to the pixel block 41) included in event data to a pixel signal of the pixel 51 at the location (x, y) and setting a predetermined color such as gray to pixels at other locations.

Note that, in the frame volume, there are a plurality of pieces of event data that are different in the event time point t but the same in the event location (x, y) in some cases. In this case, for example, event data at the latest or oldest event time point t can be prioritized. Further, in the case where event data includes polarities, the polarities of a plurality of pieces of event data that are different in the event time point t but the same in the event location (x, y) can be added together, and a pixel value based on the added value obtained by the addition can be set to a pixel at the event location (x, y).

Here, in a case where the frame width and the frame interval are the same, the frame volumes are adjacent to each other without any gap. Further, in a case where the frame interval is larger than the frame width, the frame volumes are arranged with gaps. In a case where the frame width is larger than the frame interval, the frame volumes are arranged to be partly overlapped with each other.

Fig. 9 is a block diagram illustrating another configuration example of the quantization section 84 of Fig. 5.

Note that, in Fig. 9, parts corresponding to those in the case of Fig. 7 are denoted by the same reference signs, and the description thereof is omitted as appropriate below.

In Fig. 9, the quantization section 84 includes comparators 111 and 112 and an output section 113.

Thus, the quantization section 84 of Fig. 9 is similar to the case of Fig. 7 in including the comparator 111. However, the quantization section 84 of Fig. 9 is different from the case of Fig. 7 in newly including the comparator 112 and the output section 113.

The event detecting section 52 (Fig. 5) including the quantization section 84 of Fig. 9 detects, in addition to events, the polarities of changes in light amount that are events.

In the quantization section 84 of Fig. 9, the comparator 111 outputs, in the case where a difference signal is larger than the threshold Vth, the H level indicating 1, as event data indicating the occurrence of an event having the positive polarity. The comparator 111 outputs, in the case where a difference signal is not larger than the threshold Vth, the L level indicating 0, as event data indicating that no event having the positive polarity has occurred.

Further, in the quantization section 84 of Fig. 9, a threshold Vth' (<Vth) is supplied to the non-inverting input terminal (+) of the comparator 112, and difference signals are supplied to the inverting input terminal (-) of the comparator 112 from the subtraction section 83. Here, for the sake of simple description, it is assumed that the threshold Vth' is equal to -Vth, for example, which needs however not to be the case.

The comparator 112 compares a difference signal from the subtraction section 83 with the threshold Vth' applied to the inverting input terminal (-), thereby quantizing the difference signal. The comparator 112 outputs, as event data, the quantized value obtained by the quantization.

For example, in a case where a difference signal is smaller than the threshold Vth' (the absolute value of the difference signal having a negative value is larger than the threshold Vth), the comparator 112 outputs the H level indicating 1, as event data indicating the occurrence of an event having the negative polarity. Further, in a case where a difference signal is not smaller than the threshold Vth' (the absolute value of the difference signal having a negative value is not larger than the threshold Vth), the comparator 112 outputs the L level indicating 0, as event data indicating that no event having the negative polarity has occurred.

The output section 113 outputs, on the basis of event data output from the comparators 111 and 112, event data indicating the occurrence of an event having the positive polarity, event data indicating the occurrence of an event having the negative polarity, or event data indicating that no event has occurred to the transfer section 85.

For example, the output section 113 outputs, in a case where event data from the comparator 111 is the H level indicating 1, +V volts indicating +1, as event data indicating the occurrence of an event having the positive polarity, to the transfer section 85. Further, the output section 113 outputs, in a case where event data from the comparator 112 is the H level indicating 1, -V volts indicating -1, as event data indicating the occurrence of an event having the negative polarity, to the transfer section 85. In addition, the output section 113 outputs, in a case where each event data from the comparators 111 and 112 is the L level indicating 0, 0 volts (GND level) indicating 0, as event data indicating that no event has occurred, to the transfer section 85.

The transfer section 85 supplies a request to the arbiter 33 in the case where it is confirmed on the basis of event data from the output section 113 of the quantization section 84 that a change in light amount that is an event having the positive polarity or the negative polarity has occurred. After receiving a response indicating event data output permission, the transfer section 85 outputs event data indicating the occurrence of the event having the positive polarity or the negative polarity (+V volts indicating 1 or -V volts indicating -1) to the output section 35.

Preferably, the quantization section 84 has a configuration as illustrated in Fig. 9.

Fig. 10 is a diagram illustrating another configuration example of the event detecting section 52.

In Fig. 10, the event detecting section 52 includes a subtractor 430, a quantizer 440, a memory 451, and a controller 452. The subtractor 430 and the quantizer 440 correspond to the subtraction section 83 and the quantization section 84, respectively.

Note that, in Fig. 10, the event detecting section 52 further includes blocks corresponding to the currentvoltage converting section 81 and the buffer 82, but the illustrations of the blocks are omitted in Fig. 10.

The subtractor 430 includes a capacitor 431, an operational amplifier 432, a capacitor 433, and a switch 434. The capacitor 431, the operational amplifier 432, the capacitor 433, and the switch 434 correspond to the capacitor 101, the operational amplifier 102, the capacitor 103, and the switch 104, respectively.

The quantizer 440 includes a comparator 441. The comparator 441 corresponds to the comparator 111.

The comparator 441 compares a voltage signal (difference signal) from the subtractor 430 with the predetermined threshold voltage Vth applied to the inverting input terminal (-). The comparator 441 outputs a signal indicating the comparison result, as a detection signal (quantized value).

The voltage signal from the subtractor 430 may be input to the input terminal (-) of the comparator 441, and the predetermined threshold voltage Vth may be input to the input terminal (+) of the comparator 441.

The controller 452 supplies the predetermined threshold voltage Vth applied to the inverting input terminal (-) of the comparator 441. The threshold voltage Vth which is supplied may be changed in a time-division manner. For example, the controller 452 supplies a threshold voltage Vthl corresponding to ON events (for example, positive changes in photocurrent) and a threshold voltage Vth2 corresponding to OFF events (for example, negative changes in photocurrent) at different timings to allow the single comparator to detect a plurality of types of address events (events).

The memory 451 accumulates output from the comparator 441 on the basis of Sample signals supplied from the controller 452. The memory 451 may be a sampling circuit, such as a switch, plastic, or capacitor, or a digital memory circuit, such as a latch or flip-flop. For example, the memory 451 may hold, in a period in which the threshold voltage Vth2 corresponding to OFF events is supplied to the inverting input terminal (-) of the comparator 441, the result of comparison by the comparator 441 using the threshold voltage Vthl corresponding to ON events. Note that, the memory 451 may be omitted, may be provided inside the pixel (pixel block 41), or may be provided outside the pixel.

Fig. 11 is a block diagram illustrating another configuration example of the pixel array section 31 of Fig. 2.

Note that, in Fig. 11, parts corresponding to those in the case of Fig. 3 are denoted by the same reference signs, and the description thereof is omitted as appropriate below.

In Fig. 11, the pixel array section 31 includes the plurality of pixel blocks 41. The pixel block 41 includes the lx J pixels 51 that are one or more pixels and the event detecting section 52.

Thus, the pixel array section 31 of Fig. 11 is similar to the case of Fig. 3 in that the pixel array section 31 includes the plurality of pixel blocks 41 and that the pixel block 41 includes one or more pixels 51 and the event detecting section 52. However, the pixel array section 31 of Fig. 11 is different from the case of Fig. 3 in that the pixel block 41 does not include the pixel signal generating section 53.

As described above, in the pixel array section 31 of Fig. 11, the pixel block 41 does not include the pixel signal generating section 53, so that the sensor section 21 (Fig. 2) can be formed without the AD conversion section 34.

Fig. 12 is a circuit diagram illustrating a configuration example of the pixel block 41 of Fig. 11.

As described with reference to Fig. 11, the pixel block 41 includes the pixels 51 and the event detecting section 52, but does not include the pixel signal generating section 53.

In this case, the pixel 51 can only include the photoelectric conversion element 61 without the transfer transistors 62 and 63.

Note that, in the case where the pixel 51 has the configuration illustrated in Fig. 12, the event detecting section 52 can output a voltage corresponding to a photocurrent from the pixel 51, as a pixel signal.

Fig. 13 is a block diagram illustrating a configuration example of a scan type imaging device which may be used as an EVS.

As illustrated in Fig. 13, an imaging device 510 includes a pixel array section 521, a driving section 522, a signal processing section 525, a read-out region selecting section 527, and an optional signal generating section 528.

The pixel array section 521 includes a plurality of pixels 530. The plurality of pixels 530 each output an output signal in response to a selection signal from the read-out region selecting section 527. The plurality of pixels 530 can each include an in-pixel quantizer as illustrated in Fig. 10, for example. The plurality of pixels 530 outputs output signals corresponding to the amounts of change in light intensity. The plurality of pixels 530 may be two-dimensionally disposed in a matrix as illustrated in Fig. 13.

The driving section 522 drives the plurality of pixels 530, so that the pixels 530 output pixel signals generated in the pixels 530 to the signal processing section 525 through an output line 514. Note that, the driving section 522 and the signal processing section 525 are circuit sections for acquiring grayscale information.

The read-out region selecting section 527 selects some of the plurality of pixels 530 included in the pixel array section 521. For example, the read-out region selecting section 527 selects one or a plurality of rows included in the two-dimensional matrix structure corresponding to the pixel array section 521. The readout region selecting section 527 sequentially selects one or a plurality of rows on the basis of a cycle set in advance, e.g. based on a rolling shutter. Further, the read-out region selecting section 527 may determine a selection region on the basis of requests from the pixels 530 in the pixel array section 521.

The optional signal generating section 528 may generate, on the basis of output signals of the pixels 530 selected by the read-out region selecting section 527, event signals corresponding to active pixels in which events have been detected of the selected pixels 530. The events mean an event that the intensity of light changes. The active pixels mean the pixel 530 in which the amount of change in light intensity corresponding to an output signal exceeds or falls below a threshold set in advance. For example, the signal generating section 528 compares output signals from the pixels 530 with a reference signal, and detects, as an active pixel, a pixel that outputs an output signal larger or smaller than the reference signal. The signal generating section 528 generates an event signal (event data) corresponding to the active pixel.

The signal generating section 528 can include, for example, a column selecting circuit configured to arbitrate signals input to the signal generating section 528. Further, the signal generating section 528 can output not only information regarding active pixels in which events have been detected, but also information regarding non-active pixels in which no event has been detected.

The signal generating section 528 outputs, through an output line 515, address information and timestamp information (for example, (X, Y, T)) regarding the active pixels in which the events have been detected. However, the data that is output from the signal generating section 528 may not only be the address information and the timestamp information, but also information in a frame format (for example, (0, 0, 1, o, -)).

In the following description reference will mainly be made to sensor devices of the EVS type as described above in order to ease the description and to cover an important application example. However, the principles explained below apply just as well to general imaging devices.

Fig. 14 shows a schematic illustration of a sensor device 10 that comprises a vision sensor 1010, a modulation unit 1020, and a processing unit 1030.

The vision sensor 1010 includes a pixel array 1011 that comprises a plurality of pixels 51 that are each configured to receive light and to perform photoelectric conversion to generate an electrical signal. The pixels 51 of the vision sensor 1010 may be standard imaging pixels that are configured to capture RGB or grayscale images of a scene on a frame basis. However, the pixels 51 may also be event detection pixels that are each configured to asynchronously generate event data based on the received light, which event data indicate as an event the occurrence of an intensity change of the light above an event detection threshold. Thus, the vision sensor 1010 may constitute an EVS as described above with respect to Figs. 1 to 13. The pixel array 1011 may be formed solely of event detection pixels or may be a hybrid sensor array comprising a mixture of event detection pixels and pixels that generate an electrical signal that indicates the absolute intensity of the received light. Also, the pixels 51 of the pixel array 1010 may have the ability to generate both, event data and intensity data, as was described above with respect to Figs. 3 and 4.

The vision sensor 1010 may comprise further components besides the pixel array 1011, based on which components a data stream is formed and output by the vision sensor 1010. For example, the vision sensor 1010 may comprise a classical digital signal processor that is configured to process the raw electrical signals provided by the pixels 51 in an in principle known manner. Alternatively or additionally, the vision sensor 1010 may comprise a machine learning module 1015 (termed third machine learning module 1015), preferably constituted by a neural network, that is configured to operate on the raw or pre- processed electrical signals provided by the pixels 51. The machine learning module 1015 may for example preprocess the gathered electrical signals/data instead of a digital signal processor.

Examples of machine learning algorithms that operate on event data are e.g. given in “Event-based Asynchronous Sparse Convolutional Networks” by Messikommer et al., “Unsupervised Feature Learning for Event Data: Direct vs Inverse Problem Formulation” by Kostadinov and Scaramuzza, “AEGNN: Asynchronous Event-based Graph Neural Networks” by Schaefer et al., “Event Transformer” by Li et al., and “Recurrent Vision Transformers for Object Detection with Event Cameras” by Gehrig and Scaramuzza, the content of which is incorporated herein by reference. These or similar algorithms might be used in the machine learning module 1015 of the vision sensor 1010.

The modulation unit 1020 of the sensor device 10 is configured to receive portions of the data stream and to modulate said portions of the data stream to generate modulated data portions. That is, parts of the data stream are changed by the modulation unit 1020. The term “modulation” shall indicate in this context that the change is of a nature that does not extensively alter the data contained in the data stream. In particular, the original information of the data stream will not be lost due to the modulation. Instead, important features of the data stream might be enhanced, or additional information might be embedded in the data stream without deteriorating the original information. For example, in an event data stream events may be removed or added from the stream, e.g. to pronounce certain features. In an RGB intensity signal the effects of modulation should not be perceptually noticeable in a corresponding RGB image or should improve the image.

Here, the portions of the data stream could be a temporal and/or spatial selection out of data of the data stream. For example, for an event data stream event frames might be generated as exemplarily described with respect to Fig. 8, which event frames constitute the portions of the data steam. Just the same, image frames of a conventional camera can be considered as portions of an image data stream. The data stream may also be divided spatially such that specific parts of an (event or image) frame constitute portions of the data stream. However, data portions are not necessarily continuous in space and time and may be arbitrarily selected from the data available for a given time period, if this proofs to be advantageous. The portioning of the data stream may for example be executed by the machine learning module 1015 of the vision sensor 1010.

In the modulation unit 1020 the modulation is carried out by a machine learning module 1025 (termed second machine learning module 1025), which is different from the machine learning module 1015 of the vision sensor 1010. The machine learning module 1025 of the modulation unit 1020 may in principle be an arbitrary machine learning algorithm that is capable to carry out the functions of the modulation unit 1020 described below. In particular, a neural network having an in principle known structure might be used that carries out the functions of the modulation unit 1020.

The modulation unit 1020 provides the modulated data portions to the processing unit 1030. The processing unit 1030 is configured to carry out a predetermined processing on the data stream by using a machine learning module 1035 (termed first machine learning module 1035). The predetermined processing might be any task that can be commonly applied to data of an imaging sensor, like e.g. image generation, image classification, classification of movements, object classification, object detection, and image segmentation.

That is, in principle the data stream is forwarded to the processing unit 1030 for further processing. The processing unit 1030 may carry out any kind of predetermined operation, although of particular interest are operations in the field of image processing. For example, the processing unit 1030 may operate on pixel signals/the data stream to generate an image that is free of imaging artifacts like noise or inter-pixel mismatch of the pixel array 1011 or handshake during image acquisition. The processing unit 1030 may compensate such effects. Additionally or alternatively, the processing unit 1030 may also operate on the data stream without generating an image (or without generating an image that is appealing for a human observer). For example, the processing unit 1030 may execute classification and/or detection tasks. It may classify the pixel signals/the data stream according to the observed scenes (e.g. countryside, city) or may classify and/or detect objects (e.g. persons, cars, roadsides) or movements (e.g. hand gestures, approaching objects) within the observed scenes. Further, the processing unit 1030 may also segment observed scenes (e.g. healthy tissue - pathological tissue, road - curb). All these tasks or types of predetermined processing can be executed by a machine learning module, preferably by a neural network, that has been designed in an in principle known manner for the task at hand.

The processing unit 1030 will carry out the predetermined processing based on the modulated data portions, but not on the original data stream. This provides on the one hand the possibility to optimize the data stream such that the predetermined processing achieves better results or achieves results with less computational complexity and latency. On the other hand, it is possible to ensure that the processing unit 1030 receives only data that have been changed as intended by the modulation. For example, if anti-fraud information such as an authentication certificate, an electronic signature, information about the origin of the data or the like is embedded into the data stream by the modulation, this might be used to allow the processing unit 1030 to output only results of a predetermined processing of data streams for which the authenticity can be established or to completely prohibit such processing.

Here, possible modulation schemes that can be performed by the modulation unit 1020 to modify the portions of the data stream may for example be the reduction of noise components of the data stream, the embedding of predetermined contents into the data stream, or preprocessing the data stream such as to optimize the predetermined processing.

An example for the reduction of noise components is shown in Fig. 15A. As can be seen, in this modulation scheme it is tried to delete noise component from an image. Fig. 15A shows on the left side a noisy image that is represented by grayscale or RGB intensity values, or that can be derived from event data. The respective portion of the data stream, i.e. the intensity frame or the set of events, is input to the modulation unit 1020/the second machine learning module 1025. The modulated data represent then an image with reduced noise as e.g. shown on the right of Fig. 15A. Algorithms for denoising with or without use of an artificial intelligence algorithm/a machine learning module are well known and are summarized for example in “Brief review of image denoising techniques” by Fan et al. which is incorporated by reference herein.

By the reduction of noise information that is not related to the captured scene has been removed from the data stream. This means that the processing unit 1030 does not need to extract the according data from the original data stream. This lowers the processing complexity for the processing unit 1030, since the truly captured scene is enhanced, which simplifies image recognition and classification tasks.

A simple example for the embedding of predetermined content is illustrated in Fig. 15B. Here, portions of a data stream representing the image shown on the left of Fig. 15B are input into the modulation unit 1020 together with additional content about the origin of the image, such as e.g. the data of image capturing, the type of device used for the capturing, a person who captured the image, or the like. This information is used by the second machine learning module 1025 to generate a modulated data portion that represents the image shown on the right of Fig. 15B. Here, it is intended that the two images that can be derived from the data before and after modulation appear to be the same, although additional content/information has been embedded into the data due to the modulation. Possible examples for such an embedding are watermarking and content fingerprinting, in particular active content fingerprinting, which are in principle well-known (see e.g. Voloshynovskiy et al.: “Active content fingerprinting: A marriage of digital watermarking and content fingerprinting”, or Kostadinov et al. “Active Content Fingerprinting Using Latent Data Representation, Extractor and Reconstructor”, which are incorporated by reference herein).

Thus, information for proofing the authenticity of the data stream or the right of a user of the sensor device 10 to possess the data stream (or images derived therefrom) can be provided within the data stream, however, without changing the data stream such that the result of the predetermined processing becomes deteriorated.

Here, the portions of the data stream might be modulated in an arbitrary manner as long as this improves the predetermined processing either in the quality of its output or in a reduction of the processing complexity, the latency, and/or the power consumption of the predetermined processing.

It should be noted that as in the above examples the modulation carried out by the modulation unit may preserve the dimension of the portions of the data stream, i.e. the modulation is no compression that strongly reduces the amount of data. For example, modulated images will have the same size and the same resolution. Modulated event streams will have the same temporal and spatial resolution. Also, there is no embedding of feature vectors of the data stream into lower dimensional embedding spaces. The aim of the concurrent modulation and predetermined processing is to modify the original data as little as possible while improving the quality and/or the result of the predetermined processing.

In this manner it is possible to improve the processing results on the one hand by easing the processing and on the other hand by enhancing the security against processing fraudulently obtained data.

The modulation scheme to be used and/or the parameters of the used modulation scheme may be determined by the modulation unit 1020 based on the type of data in the data stream and/or based on the predetermined processing. That is, the second machine learning module 1025 is informed about the predetermined processing/the desired outcome of the respective task. This means that also the content to be embedded, i.e. the modulation distortion is task dependent. At the same time, the result of the task will depend on the specific portion of the data stream that is processed. Thus, via the need to optimize the result of the task, also the modulation will depend on the specific data within the data stream. Accordingly, the change due to the modulation, i.e. the modulation distortion, will be correlated with the processed data as well as with the result of the predetermined processing. This is an easy manner to distinguish the modulation carried out by the modulation unit 1020 from common data message hiding/embedding schemes, where modulation content is used that is uncorrelated with, i.e. “orthogonal” to the data that is modulated.

As stated above, in principle, the efficiency of the predetermined processing will be enhanced by using the fist machine learning module 1035 together with the second machine learning modules 1025. However, whether the efficiency is truly enhanced will depend on the capability of the modulation unit 1020 to modulate the data portions in a manner that is useful for the processing unit 1030. This again depends on the particular task, i.e. the particular predetermined processing to be carried out by the processing unit 1030.

Thus, to improve the efficiency of the predetermined processing the first machine learning module 1035 is trained such as to optimize the predetermined processing together with the second machine learning module 1025 used by the modulation unit 1020 (and, if necessary, together with the third machine learning module 1015 of the vision sensor 1010 or any further machine learning modules). In particular, by training the machine learning modules 1025, 1035 together it can be ensured that the parameters of the second machine learning module 1025 are adjusted such that the modulated data portions will allow the processing unit 1030/the first machine learning module 1035 to give a correct estimate for the task at hand. The common training of the machine learning modules 1025, 1035 serves therefore the purpose of allowing an implicit training of the upstream second machine learning module 1025 which optimizes this module such that the downstream first machine learning module 1035 is optimized, too. This would not be possible for separately trained machine learning modules 1025, 1035.

Particular training scenarios will be described in the following with respect to a combination of the first and second machine learning modules 1035, 1025, i.e. of the machine learning modules of the processing unit 1030 and the modulation unit 1020, respectively. The below description generalizes in a straightforward manner to the inclusion of the third machine learning module 1015 of the vision sensor 1010 as well as to the inclusion of further machine learning modules. The description is restricted to two modules only to simplify the description, but not to limit the present disclosure. Moreover, in the following it can be assumed that the machine learning modules are constituted by neural networks or by other parametric or nonparametric models.

The machine learning modules may be trained according to two variants. Either, the modules are pretrained during a simulation stage, i.e. a stage where the operation of the sensor device 10 is simulated based on a training data set that contains a plurality of videos representing observable scenes. Based on these videos outputs of the vision sensor 1010 are simulated which are used to simulate the performance of the first and second machine learning modules 1035, 1025. The performance of the machine learning modules 1025, 1035 is optimized and the parameters of the optimized models are fixedly stored in the modulation unit 1020 and the processing unit 1030.

Or, the models are trained based on a continual learning algorithm. Then, each newly observed scene will lead to a re-evaluation of the machine learning modules 1025, 1035 that might lead to an adaption of the parameters of the modules.

Further, supervised and un-supervised learning may be used for the training, where supervised learning is preferably used in the pre-training variant, while un-supervised learning might be applied in both variants.

Feedback paths from the output of the processing unit that are used during the training process are illustrated with broken lines in Fig. 14. However, it should be noted that these feedback paths only represent the conceptual flow of information but not necessarily the actual transfer of data.

In the pre-training setup, a joint learning objective or loss function can be defined for both machine learning modules:

L = M L + p Lp by adding a loss function LM for the machine learning module 1025 of the modulation unit 1020 and a loss function Lp for the machine learning module 1035 of the processing unit 1030 which are weighted with tunable parameters A\i and Xp. The parameters of the corresponding neural networks can then be optimized by using backpropagation, preferably with gradient descent or more preferably with stochastic gradient descent as e.g. described in “Deep Learning, volume 1” by Goodfellow et al. (MIT Press, 2016), the content of which his hereby incorporated by reference. Here, the Adam optimization algorithm can be used (see e.g. “Adam: A Method for Stochastic Optimization” by Kingma and Ba, the content of which is hereby incorporated by reference). Also, first or second order gradient-based methods may be used.

Training labels, i.e. desired estimations of the predetermined processing, may be only provided for the combined task, i.e. for the output of the first machine learning module 1035, while labels for the loss function of the second machine learning module may be deducible therefrom or may be defined manually. This is in particular useful for cases, where the modulation unit 1020 is used as a pre-processor that optimizes the input for the first machine learning module 1035, however, without knowing in advance in which manner to modulate the portions of the data stream. In this case, labels for the modulation could be deduced by simulating various modulations and looking for improvements of the result of the predetermined processing

For a known manner of modulation, i.e. a modulation type that is not chosen by the second machine learning module 1025, also other criteria could be used to explicitly define the modulation loss LM. For example, if denoising is chosen, quality metrics known from denoising algorithms might be chosen, such as peak signal-to-noise ratio, PSNR, metrics, structural similarity index metrics, SSIM, or mean square error, MSE, metrics, and training labels can be set based on these metrics. For embedding modulations an objective might be to minimize the modulation distortion between the modulated and the original data stream. Also here in principle known metrics like MSE metrics might be used to determine training labels.

Moreover, by backpropagating the entire/combined loss L over the combined neural network, it will be possible to optimize the network even without using exact expression for LM and Lp, since the error of the final outcome can be used to adjust weights of the nodes of the combined neural network.

In the supervised, pre-training setup the loss function Lp for the first machine learning module 1035 may be based on a similarity metric and can be cross entropy loss or mean square error loss and the loss function L for the second machine learning module 1025 may also be based on a similarity metric and can be cross entropy loss, connectionist temporal classification, CTC, loss, or mean square error loss. However, any other loss functions may be used that can efficiently help to optimize the predetermined processing.

In a continual learning setup the same principle formula for the total loss function can be used. In this case backpropagation as used in the pre-training setup may be combined with experience repay as e.g. explained in “Learning and Categorization in Modular Neural Networks” by Murre or other methods that avoid catastrophic forgetting and writing of neural network parameters as e.g. described in “Catastrophic Forgetting in Connectionist Networks” by French, which documents are both incorporated by reference herein. Of course, any other well-known continual learning algorithm may be used. In the self-supervised/unsupervised learning case there will be no labels for the task at hand, i.e. for the output of the predetermined processing. Nevertheless, it is possible to define labels for the modulation loss LM using the loss Lp of the first machine learning module 1035. Here, the loss function Lp of the first machine learning module 1035 may be based on a similarity metric and may be mean square error loss or contrastive loss (see e.g. “A Simple Framework for Contrastive Learning of Visual Representations” by Chen, which is incorporated by reference herein) and the loss function LM of the second machine learning module 1025 may be based on a similarity metric and may be cross entropy loss (see e.g. “Deep Learning, volume 1” by Goodfellow et al, which is incorporated by reference herein).

The second machine learning module 1025 may be a recurrent neural network or a long short-term memory, LSTM, network. Thus, the second machine learning module 1025 may be considered a neural network with a memory unit that allows an operation on single events. Due to the capability to store information for several time steps, such networks are able to understand the temporal correlation between different events. Nevertheless, they can work with the high temporal resolution of single events in the order of microseconds. On the other hand, also neural networks without memory units might be used. Although such networks are simpler, data have to be pre-processed, if temporal structure are to be observable by them. In particular, events need to be accumulated/integrated, for example into event frames, if neural networks without memory units are to be used.

As should have become clear from the above description, in principle known ways of training neural networks/machine learning modules can be combined to provide an improved integration of neural networks that increases the efficiency with which tasks in image processing can be executed.

In the above description the processing unit 1030 may be constituted by a commonly known device and may for example be a computer, a processor, a CPU, a GPU, circuitry, software, a program or application running on a processor and the like. Moreover, the computational functions of the modulation unit 1020 and the vision sensor 1010 may also be carried out by any known device such a processor, a CPU, a GPU, circuitry, software, a program or application on a processor and the like. Thus, the nature of the computing devices executing the functions of the vision sensor 1010, the modulation unit 1020 and the processing unit 1030 are arbitrary, as long as they are configured to carry out the functions described herein.

As illustrated in Fig. 14 the vision sensor 1010 and the modulation unit 1020 may be formed on the same sensor chip 11, while the processing unit 1030 is on a separate chip/die/substrate or part of an external device. However, the modulation unit 1020 may also be part of the chip carrying the processing unit 1030 Moreover, vision sensor 1010, modulation unit 1020 and processing unit 1030 may be arranged on the same chip or may each be located on different chips.

The above-described method for operating a sensor device is summarized again in Fig. 16. At S 101, by a vision sensor 1010 that comprises a pixel array 1011 having a plurality of pixels 51 each being configured to receive light and to perform photoelectric conversion to generate an electrical signal, a data stream is formed and output based on the generated electrical signals. At SI 02 portions of the data stream are received by a modulation unit 1020 that modulates said portions of the data stream to generate modulated data portions by using a second machine learning module.

At SI 03 the modulated data portions are received by a processing unit 1040 that carries out a predetermined processing of the modulated data portions by using a first machine learning module 1035.

At SI 04 the first machine learning module 1035 is trained such as to optimize the predetermined processing together with a second machine learning module 1025.

Here, it should be noted that although Fig. 16 shows different steps that follow each other subsequently, this order is not necessarily fixed. In particular, training of the first machine learning module 1035 may precede the formation and output of the data stream in a pre-training setup or may run concurrent with the predetermined processing.

As explained above, pre-processing of the pixel signals provided by the pixels 51 may be necessary in particular for event detection pixels. In the following, based on Fig. 17 possible modulation setups that are based on an event stream will be discussed that are based on event frames. The generation of event frames is typically done within the vision sensor 1010 and may or may not be carried out by using the third machine learning module 1015.

The grouping of events into event frames F may be done as discussed above with respect to Fig. 8. In this example, temporal planes are defined that indicate the end and the beginning of one event frame, and events lying between these planes are used to generate pixel values of the event frame, e.g. by adding 1 for a positive polarity event in a pixel 51 and -1 for a negative polarity event. However, event frames might also be generated differently. For example, positive and negative event values may be projected in a weighted manner to temporal planes defining the position of the event frames, where weights depend on the temporal distance to the respective plane. Pixel values are then obtained by adding all weights obtained for one pixel.

Fig. 17 shows schematically this division of the event stream generated by the pixel array 1011 into event frames F. As shown in Fig. 17 the event stream extends two-dimensionally across the coordinates of the pixel array 1011 (x and y coordinates). The event stream is sliced in the time direction t. Single event frames F constitute portions of the event stream that are input into the modulation unit 1020. The modulation unit 1020 modulates each event frame F separately and provides the modulated event frames F’ to the processing unit 1030 The processing unit 1030 will carry out the predetermined processing on the modulated event frame F’ and will produce a processing result. This processing result can be evaluated as described above with an appropriate metric/loss function. The results of this evaluation are fed back to the first and second machine learning modules 1035, 1025 for training of the modules either in a separate pre-training step or for continual learning as discussed above. Here, also groups of event frames F may be processed/modulated at once. In this manner, event frames F can be used as a simple manner of ordering the event stream generated by the pixel array 1011 such as to allow meaningful and resource-efficient processing of the event data. Of course, also other manners of dividing and ordering an event stream are possible. For example, voxel grid representations can be chosen, or event frames could be further spatially divided to generate event subframes, where each subframe is modulated separately. Finally, it is also possible to use the raw event data for modulation and processing.

As described above, a possible manner of modulation is the reduction of noise contained within the event stream/the event frames. In this situation the modulation loss function could in particular be a mean square error loss that compares the modulated frames F’ with a desired output/label. The labels can be provided externally. For example, during pre-training pairs of data sets can be used, the one containing noise and the other containing no noise. While the noise data set will be used to simulate the behavior of the sensor device, the noise-less data set can be used as ground-truth. However, labels could also be determined based on the predetermined processing, i.e. such that the de-noising makes the data fit for the processing. The loss function of the predetermined processing can be chosen according to the processing at hand. If e.g. image reconstruction from the events is the task, then also in this case a mean square error loss can be used that compares the result of the image reconstruction that is based on the modulated event frames with a ground-truth, noiseless representation of the corresponding scene.

A second main example is the embedding of content into the original event frames. As illustrated in Fig. 17, in this case also the content to be added is provided to the modulation unit 1020, which uses an appropriate modulation scheme to hide this content in the modulated event frames F’. The reconstruction of this content from the modulated event frames F’ is then a possible alternative or additional task of the processing unit 1030/the first machine learning module. One example for content to be hidden is constituted by active content fingerprinting as e.g. described in Voloshynovskiy et al.: “Active content fingerprinting: A marriage of digital watermarking and content fingerprinting”, or Kostadinov et al. “Active Content Fingerprinting Using Latent Data Representation, Extractor and Reconstructor”, which are incorporated by reference herein.

The loss function of the modulation unit 1020/the second machine learning module 1025 for content hiding may take into account that the modulation distortion should be minimized. On the other hand, each modulation should improve the outcome of the predetermined processing, i.e. the processing result. Thus, as a loss function for the second machine learning module 1025 mean square error loss can be used that compares the original event frames F with the modulated event frames F’. The loss function of the first machine learning module will have two components. The first being related to the optimization of the predetermined processing, i.e. image reconstruction, object recognition etc.. The second component will indicate how well the embedded content could be retrieved. Any suitable and in principle well-known loss functions can be used for these two components.

All the above-described examples allow an improvement of the data stream by modulation that can be used to optimize a subsequent predetermined processing step. In this manner, various predetermined tasks can be carried out by the processing unit 1030 in a satisfactory manner. Further, it is at the same time possible to embed additional contents into the data stream that can be used to supplement the data stream with metadata or with authentication information.

The technology according to the above (i.e. the present technology) is applicable to various products. For example, the technology according to the present disclosure may be realized as a device that is installed on any kind of moving bodies, for example, vehicles, electric vehicles, hybrid electric vehicles, motorcycles, bicycles, personal mobilities, airplanes, drones, ships, and robots.

Fig. 18 is a block diagram depicting an example of schematic configuration of a vehicle control system as an example of a mobile body control system to which the technology according to an embodiment of the present disclosure can be applied.

The vehicle control system 12000 includes a plurality of electronic control units connected to each other via a communication network 12001. In the example depicted in Fig. 18, the vehicle control system 12000 includes a driving system control unit 12010, a body system control unit 12020, an outside -vehicle information detecting unit 12030, an in-vehicle information detecting unit 12040, and an integrated control unit 12050. In addition, a microcomputer 12051, a sound/image output section 12052, and a vehicle-mounted network interface (I/F) 12053 are illustrated as a functional configuration of the integrated control unit 12050.

The driving system control unit 12010 controls the operation of devices related to the driving system of the vehicle in accordance with various kinds of programs. For example, the driving system control unit 12010 functions as a control device for a driving force generating device for generating the driving force of the vehicle, such as an internal combustion engine, a driving motor, or the like, a driving force transmitting mechanism for transmitting the driving force to wheels, a steering mechanism for adjusting the steering angle of the vehicle, a braking device for generating the braking force of the vehicle, and the like.

The body system control unit 12020 controls the operation of various kinds of devices provided to a vehicle body in accordance with various kinds of programs. For example, the body system control unit 12020 functions as a control device for a keyless entry system, a smart key system, a power window device, or various kinds of lamps such as a headlamp, a backup lamp, a brake lamp, a turn signal, a fog lamp, or the like. In this case, radio waves transmitted from a mobile device as an alternative to a key or signals of various kinds of switches can be input to the body system control unit 12020. The body system control unit 12020 receives these input radio waves or signals, and controls a door lock device, the power window device, the lamps, or the like of the vehicle.

The outside-vehicle information detecting unit 12030 detects information about the outside of the vehicle including the vehicle control system 12000. For example, the outside-vehicle information detecting unit 12030 is connected with an imaging section 12031. The outside-vehicle information detecting unit 12030 makes the imaging section 12031 image an image of the outside of the vehicle, and receives the imaged image. On the basis of the received image, the outside-vehicle information detecting unit 12030 may perform processing of detecting an object such as a human, a vehicle, an obstacle, a sign, a character on a road surface, or the like, or processing of detecting a distance thereto.

The imaging section 12031 is an optical sensor that receives light, and which outputs an electric signal corresponding to a received light amount of the light. The imaging section 12031 can output the electric signal as an image, or can output the electric signal as information about a measured distance. In addition, the light received by the imaging section 12031 may be visible light, or may be invisible light such as infrared rays or the like.

The in-vehicle information detecting unit 12040 detects information about the inside of the vehicle. The in-vehicle information detecting unit 12040 is, for example, connected with a driver state detecting section 12041 that detects the state of a driver. The driver state detecting section 12041, for example, includes a camera that images the driver. On the basis of detection information input from the driver state detecting section 12041, the in-vehicle information detecting unit 12040 may calculate a degree of fatigue of the driver or a degree of concentration of the driver, or may determine whether the driver is dozing.

The microcomputer 12051 can calculate a control target value for the driving force generating device, the steering mechanism, or the braking device on the basis of the information about the inside or outside of the vehicle which information is obtained by the outside-vehicle information detecting unit 12030 or the in-vehicle information detecting unit 12040, and output a control command to the driving system control unit 12010. For example, the microcomputer 12051 can perform cooperative control intended to implement functions of an advanced driver assistance system (ADAS) which functions include collision avoidance or shock mitigation for the vehicle, following driving based on a following distance, vehicle speed maintaining driving, a warning of collision of the vehicle, a warning of deviation of the vehicle from a lane, or the like.

In addition, the microcomputer 12051 can perform cooperative control intended for automatic driving, which makes the vehicle to travel autonomously without depending on the operation of the driver, or the like, by controlling the driving force generating device, the steering mechanism, the braking device, or the like on the basis of the information about the outside or inside of the vehicle which information is obtained by the outside-vehicle information detecting unit 12030 or the in-vehicle information detecting unit 12040.

In addition, the microcomputer 12051 can output a control command to the body system control unit 12020 on the basis of the information about the outside of the vehicle which information is obtained by the outside-vehicle information detecting unit 12030. For example, the microcomputer 12051 can perform cooperative control intended to prevent a glare by controlling the headlamp so as to change from a high beam to a low beam, for example, in accordance with the position of a preceding vehicle or an oncoming vehicle detected by the outside-vehicle information detecting unit 12030.

The sound/image output section 12052 transmits an output signal of at least one of a sound and an image to an output device capable of visually or auditorily notifying information to an occupant of the vehicle or the outside of the vehicle. In the example of Fig. 18, an audio speaker 12061, a display section 12062, and an instrument panel 12063 are illustrated as the output device. The display section 12062 may, for example, include at least one of an on-board display and a head-up display.

Fig. 19 is a diagram depicting an example of the installation position of the imaging section 12031.

In Fig. 19, the imaging section 12031 includes imaging sections 12101, 12102, 12103, 12104, and 12105.

The imaging sections 12101, 12102, 12103, 12104, and 12105 are, for example, disposed at positions on a front nose, sideview mirrors, a rear bumper, and a back door of the vehicle 12100 as well as a position on an upper portion of a windshield within the interior of the vehicle. The imaging section 12101 provided to the front nose and the imaging section 12105 provided to the upper portion of the windshield within the interior of the vehicle obtain mainly an image of the front of the vehicle 12100. The imaging sections 12102 and 12103 provided to the sideview mirrors obtain mainly an image of the sides of the vehicle 12100. The imaging section 12104 provided to the rear bumper or the back door obtains mainly an image of the rear of the vehicle 12100. The imaging section 12105 provided to the upper portion of the windshield within the interior of the vehicle is used mainly to detect a preceding vehicle, a pedestrian, an obstacle, a signal, a traffic sign, a lane, or the like.

Incidentally, Fig. 19 depicts an example of photographing ranges of the imaging sections 12101 to 12104. An imaging range 12111 represents the imaging range of the imaging section 12101 provided to the front nose. Imaging ranges 12112 and 12113 respectively represent the imaging ranges of the imaging sections 12102 and 12103 provided to the sideview mirrors. An imaging range 12114 represents the imaging range of the imaging section 12104 provided to the rear bumper or the back door. A bird’s-eye image of the vehicle 12100 as viewed from above is obtained by superimposing image data imaged by the imaging sections 12101 to 12104, for example.

At least one of the imaging sections 12101 to 12104 may have a function of obtaining distance information. For example, at least one of the imaging sections 12101 to 12104 may be a stereo camera constituted of a plurality of imaging elements, or may be an imaging element having pixels for phase difference detection.

For example, the microcomputer 12051 can determine a distance to each three-dimensional object within the imaging ranges 12111 to 12114 and a temporal change in the distance (relative speed with respect to the vehicle 12100) on the basis of the distance information obtained from the imaging sections 12101 to 12104, and thereby extract, as a preceding vehicle, a nearest three-dimensional object in particular that is present on a traveling path of the vehicle 12100 and which travels in substantially the same direction as the vehicle 12100 at a predetermined speed (for example, equal to or more than 0 km/hour). Further, the microcomputer 12051 can set a following distance to be maintained in front of a preceding vehicle in advance, and perform automatic brake control (including following stop control), automatic acceleration control (including following start control), or the like. It is thus possible to perform cooperative control intended for automatic driving that makes the vehicle travel autonomously without depending on the operation of the driver or the like.

For example, the microcomputer 12051 can classify three-dimensional object data on three-dimensional objects into three-dimensional object data of a two-wheeled vehicle, a standard-sized vehicle, a largesized vehicle, a pedestrian, a utility pole, and other three-dimensional objects on the basis of the distance information obtained from the imaging sections 12101 to 12104, extract the classified three-dimensional object data, and use the extracted three-dimensional object data for automatic avoidance of an obstacle. For example, the microcomputer 12051 identifies obstacles around the vehicle 12100 as obstacles that the driver of the vehicle 12100 can recognize visually and obstacles that are difficult for the driver of the vehicle 12100 to recognize visually. Then, the microcomputer 12051 determines a collision risk indicating a risk of collision with each obstacle. In a situation in which the collision risk is equal to or higher than a set value and there is thus a possibility of collision, the microcomputer 12051 outputs a warning to the driver via the audio speaker 12061 or the display section 12062, and performs forced deceleration or avoidance steering via the driving system control unit 12010. The microcomputer 12051 can thereby assist in driving to avoid collision.

At least one of the imaging sections 12101 to 12104 may be an infrared camera that detects infrared rays. The microcomputer 12051 can, for example, recognize a pedestrian by determining whether or not there is a pedestrian in imaged images of the imaging sections 12101 to 12104. Such recognition of a pedestrian is, for example, performed by a procedure of extracting characteristic points in the imaged images of the imaging sections 12101 to 12104 as infrared cameras and a procedure of determining whether or not it is the pedestrian by performing pattern matching processing on a series of characteristic points representing the contour of the object. When the microcomputer 12051 determines that there is a pedestrian in the imaged images of the imaging sections 12101 to 12104, and thus recognizes the pedestrian, the sound/image output section 12052 controls the display section 12062 so that a square contour line for emphasis is displayed so as to be superimposed on the recognized pedestrian. The sound/image output section 12052 may also control the display section 12062 so that an icon or the like representing the pedestrian is displayed at a desired position.

An example of the vehicle control system to which the technology according to the present disclosure is applicable has been described above. The technology according to the present disclosure is applicable to the imaging section 12031 among the above-mentioned configurations. Specifically, the sensor device 10 is applicable to the imaging section 12031. The imaging section 12031 to which the technology according to the present disclosure has been applied flexibly acquires event data and performs data processing on the event data, thereby being capable of providing appropriate driving assistance.

Further possible implementations of the sensor device 10 are mobile devices 3000 such as cell phones, tablets, smart watches and the like as shown in Fig. 20A or head-mounted displays 4000 as shown in Fig. 20B. Further, the sensor device 10 is useable in augmented and/or virtual reality applications/cameras or in surveillance systems like 360° cameras. Note that, the embodiments of the present technology are not limited to the above-mentioned embodiment, and various modifications can be made without departing from the gist of the present technology.

Further, the effects described herein are only exemplary and not limited, and other effects may be provided.

Note that, the present technology can also take the following configurations.

[1] A sensor device (10) comprising: a vision sensor (1010) that comprises a pixel array (1011) having a plurality of pixels (51) each being configured to receive light and to perform photoelectric conversion to generate an electrical signal, based on which electrical signals a data stream is formed and output by the vision sensor (1010); a modulation unit (1020) that is configured to receive portions of the data stream and to modulate said portions of the data stream to generate modulated data portions; and a processing unit (1030) that is configured to receive the modulated data portions and to carry out a predetermined processing of the modulated data portions; wherein the processing unit (1030) is configured to use a first machine learning module (1035) to carry out the predetermined processing which first machine learning module (1035) has been trained such as to optimize the predetermined processing together with a second machine learning module (1025) used by the modulation unit (1020) to generate the modulated data portions.

[2] The sensor device (10) according to [1], comprising a single sensor chip (11) on which the vision sensor (1010) and the modulation unit (1020) are formed.

[3] The sensor device (10) according to any one of [1] to [2], wherein the pixels (51) are event detection pixels that are each configured to generate event data based on the received light, which event data indicate as an event the occurrence of an intensity change of the light above an event detection threshold.

[4] The sensor device (10) according to any one of [1] to [3], wherein the predetermined processing is one of image generation, image classification, classification of movements, object classification, object detection, and image segmentation.

[5] The sensor device (10) according to any one of [1] to [4], wherein the modulation unit (1020) is configured to modulate the portions of the data stream based on the type of data in the data stream and/or the predetermined processing.

[6] The sensor device (10) according to any one of [1] to [5], wherein the modulation unit (1020) is configured to modify the portions of the data stream due to the modulation by at least one of reducing a noise component of the data stream, embedding a predetermined content into the data stream, and preprocessing the data stream such as to optimize the predetermined processing.

[7] The sensor device (10) according to any one of [1] to [6], wherein the modulation carried out by the modulation unit (1020) preserves the dimension of the portions of the data stream.

[8] The sensor device (10) according to any one of [1] to [7], wherein the first machine learning module (1035) and the second machine learning module (1025) are neural networks with fixed weights that have been optimized in the common training.

[9] The sensor device (10) according to any one of [1] to [7], wherein the first machine learning module (1035) and the second machine learning module (1025) are neural networks whose weights are continuously adapted by a continual learning algorithm that optimizes the predetermined processing.

[10] The sensor device (10) according to any one of [1] to [9], wherein the second machine learning module (1025) is a recurrent neural network, RNN, or a long short-term memory, LSTM, network.

[11] A method for operating a sensor device (10), the method comprising by a vision sensor (1010) that comprises a pixel array (1011) having a plurality of pixels (51) each being configured to receive light and to perform photoelectric conversion to generate an electrical signal, forming and outputting a data stream based on the generated electrical signals; by a modulation unit (1020), receiving portions of the data stream and modulating said portions of the data stream to generate modulated data portions by using a second machine learning module (1025); by a processing unit (1030), receiving the modulated data portions and carrying out a predetermined processing of the modulated data portions by using a first machine learning module (1035); and training the first machine learning module (1035) together with the second machine learning module (1025) such as to optimize the predetermined processing.

[ 12] The method according to [ 11 ] , wherein training comprises defining a loss function for each of the machine learning modules that are trained together, and jointly optimizing the loss functions of the machine learning modules that are trained together.

[13] The method according to [12], wherein if training is performed by applying supervised learning, the loss function for the first machine learning module is based on a similarity metric, and is preferably cross entropy loss or mean square error loss and the loss function for the second machine learning module is based on a similarity metric and is preferably, mean square error loss, cross entropy loss or connectionist temporal classification, CTC, loss; and if training is performed by applying unsupervised learning the loss function of the first machine learning module (1035) is based on a similarity metric and is preferably mean square error loss or contrastive loss and the loss function of the second machine learning module (1025) is based on a similarity metric and is preferably cross entropy loss.

[14] The method according to any one of [11] to [13], wherein training is carried out by applying a backpropagation algorithm with, preferably stochastic, gradient descent, more preferably by using the Adam algorithm or first or second order gradient based methods.

[15] The method according to any one of [11] to [13], wherein training is carried out by applying a continual learning algorithm.

Claims

1. A sensor device comprising: a vision sensor that comprises a pixel array having a plurality of pixels each being configured to receive light and to perform photoelectric conversion to generate an electrical signal, based on which electrical signals a data stream is formed and output by the vision sensor; a modulation unit that is configured to receive portions of the data stream and to modulate said portions of the data stream to generate modulated data portions; and a processing unit that is configured to receive the modulated data portions and to carry out a predetermined processing of the modulated data portions; wherein the processing unit is configured to use a first machine learning module to carry out the predetermined processing which first machine learning module has been trained such as to optimize the predetermined processing together with a second machine learning module used by the modulation unit to generate the modulated data portions.

2. The sensor device according to claim 1, comprising a single sensor chip on which the vision sensor and the modulation unit are formed.

3. The sensor device according to claim 1, wherein the pixels are event detection pixels that are each configured to generate event data based on the received light, which event data indicate as an event the occurrence of an intensity change of the light above an event detection threshold.

4. The sensor device according to claim 1, wherein the predetermined processing is one of image generation, image classification, classification of movements, object classification, object detection, and image segmentation.

5. The sensor device according to claim 1, wherein the modulation unit is configured to modulate the portions of the data stream based on the type of data in the data stream and/or the predetermined processing.

6. The sensor device according to claim 1, wherein the modulation unit is configured to modify the portions of the data stream due to the modulation by at least one of reducing a noise component of the data stream, embedding a predetermined content into the data stream, and preprocessing the data stream such as to optimize the predetermined processing.

7. The sensor device according to claim 1, wherein the modulation carried out by the modulation unit preserves the dimension of the portions of the data stream.

8. The sensor device according to claim 1, wherein the first machine learning module and the second machine learning module are neural networks with fixed weights that have been optimized in the common training.

9. The sensor device according to claim 1, wherein the first machine learning module and the second machine learning module are neural networks whose weights are continuously adapted by a continual learning algorithm that optimizes the predetermined processing.

10. The sensor device according to claim 1, wherein the second machine learning module is a recurrent neural network, RNN, or a long shortterm memory, LSTM, network.

11. A method for operating a sensor device, the method comprising by a vision sensor that comprises a pixel array having a plurality of pixels each being configured to receive light and to perform photoelectric conversion to generate an electrical signal, forming and outputting a data stream based on the generated electrical signals; by a modulation unit, receiving portions of the data stream and modulating said portions of the data stream to generate modulated data portions by using a second machine learning module; by a processing unit, receiving the modulated data portions and carrying out a predetermined processing of the modulated data portions by using a first machine learning module; and training the first machine learning module together with the second machine learning module such as to optimize the predetermined processing.

12. The method according to claim 11, wherein training comprises defining a loss function for each of the machine learning modules that are trained together, and jointly optimizing the loss functions of the machine learning modules that are trained together.

13. The method according to claim 12, wherein if training is performed by applying supervised learning, the loss function for the first machine learning module is based on a similarity metric, and is preferably cross entropy loss or mean square error loss and the loss function for the second machine learning module is based on a similarity metric and is preferably, mean square error loss, cross entropy loss or connectionist temporal classification, CTC, loss; and if training is performed by applying unsupervised learning the loss function of the first machine learning module is based on a similarity metric and is preferably mean square error loss or contrastive loss and the loss function of the second machine learning module is based on a similarity metric and is preferably cross entropy loss.

14. The method according to claim 11, wherein training is carried out by applying a backpropagation algorithm with, preferably stochastic, gradient descent, more preferably by using the Adam algorithm or first or second order gradient based methods.

15. The method according to claim 11, wherein training is carried out by applying a continual learning algorithm.