Data processing method and system for single-photon fog penetrating imaging

Abstract

The application provides a data processing method and system for single-photon fog imaging. The method comprises: collecting, by an airborne terminal, original echo signals of a single-photon detection system for a detection target, and converging the original echo signals to obtain a plurality of high-frequency pulse data packets; performing anti-jitter buffering and dynamic load reorganization on the high-frequency pulse data packets to obtain target large packet data, and enqueuing the target large packet data to a pre-set asynchronous ring queue; transferring the target large packet data in the asynchronous ring queue to a ground terminal according to a pre-set sending thread; constructing, by the ground terminal, a three-dimensional photon counting matrix including time depth and spatial dimensions for one or more target large packet data received; and inputting the three-dimensional photon counting matrix into a pre-trained depth reconstruction network to obtain a three-dimensional depth distance map of the detection target. The technical scheme provided by the application can improve the stability of data transmission of the airborne terminal and the accuracy of the depth image generated by the ground terminal.

Description

Technical Field

[0001] This invention relates to the field of photoelectric detection technology, and in particular to a data processing method and system for single-photon fog-penetrating imaging. Background Technology

[0002] Single-photon lidar, especially detection systems based on Geiger-mode avalanche photodiodes (GM-APDs), possesses extremely high detection sensitivity at the single-photon level, making it valuable for long-range 3D imaging and fine target inspection by unmanned aerial vehicles (UAVs). This system constructs a 3D spatiotemporal photon count histogram by emitting laser pulses towards the target and accumulating multiple echo events, thereby extracting depth and distance information from the target surface.

[0003] In typical UAV field operations, the high-frequency pulse data collected by the airborne single-photon detection system needs to be transmitted to the ground for processing and imaging in real time. Conventional wireless transmission schemes typically use an airborne wireless bridge to directly forward the raw pulse data packets generated by the detector to the ground, in order to obtain complete photon counting timing information at the ground.

[0004] However, in actual complex weather conditions and UAV flight operation scenarios, the above-mentioned conventional solutions face the following two major technical challenges.

[0005] First, there is the bottleneck of wireless transmission congestion for massive high-frequency pulse data at the front end. To support the back end's precise reconstruction of weak signals under complex weather conditions, single-photon detectors must send massive amounts of fragmented raw pulse data packets to the ground in an extremely short time. These data packets are usually encapsulated in User Datagram Protocol (UDP) and have extremely strong burstiness and high-frequency characteristics. When these data packets are transmitted through a restricted airborne wireless bridge, the limited hardware buffer capacity inside the bridge makes it prone to buffer overflow, network congestion, and severe packet loss when faced with the instantaneous data surge generated by the detector. Once the original time sequence of photon counting is damaged at the transport layer, its physical distribution characteristics will be irreversibly destroyed, causing the subsequent deep reconstruction logic based on spatial priors and statistical characteristics to completely fail.

[0006] Second, there are challenges in signal distortion and target extraction under complex fog and haze environments. In clear, fog-free environments, target reflection signals exhibit a clear, single Gaussian distribution with no backscattering peaks. However, in environments with high aerosol concentrations, such as fog and haze, laser pulses interact with numerous aerosol particles along their transmission path, generating strong backscattering. This backscattering manifests as a dominant gamma-shaped scattering peak in the photon counting histogram, severely masking or even obscuring the weak Gaussian target echo. Under these extremely low signal-to-noise ratio conditions, traditional depth extraction algorithms based on local peak detection struggle to accurately distinguish between target reflection signals and scattering noise, easily leading to severe false alarms in ranging and resulting in unreliable or even completely invalid 3D imaging results.

[0007] To address the aforementioned transmission congestion issues, a direct approach is to increase the hardware buffer capacity of the wireless transmission link or improve the transmission bandwidth. However, due to the strict constraints on the size, weight, and power consumption of UAV payloads, solutions relying solely on hardware upgrades are difficult to implement on airborne platforms. Regarding the aforementioned fog and haze scattering noise issues, conventional image dehazing or peak detection algorithms can only perform local processing in the spatial or temporal domains, lacking the ability to effectively utilize global correlation information between nodes across the entire image in both time and space, and thus failing to robustly recover the target's 3D depth information from data with extremely low signal-to-noise ratios. Summary of the Invention

[0008] The data processing method and system for single-photon fog-penetrating imaging provided by the embodiments of the present invention can improve the stability of airborne data transmission and the accuracy of depth images generated on the ground to a certain extent.

[0009] In a first aspect, the present invention provides a data processing method for single-photon fog-penetrating imaging, the method being applied to a data processing system including airborne and ground-based terminals, the method comprising: The onboard terminal acquires the raw echo signal of the single-photon detection system to the target and aggregates the raw echo signal to obtain multiple high-frequency pulse data packets. The high-frequency pulse data packets are subjected to anti-jitter buffering and dynamic load reassembly to obtain the target large packet data, which is then enqueued into a pre-set asynchronous circular queue. The target large packet data in the asynchronous circular queue is transmitted to the ground terminal according to the preset sending thread; Using the ground terminal, a three-dimensional photon counting matrix, including temporal depth and spatial dimensions, is constructed for one or more target large data packets received. The three-dimensional photon counting matrix is ​​input into a pre-trained deep reconstruction network to obtain a three-dimensional depth distance map of the target being detected; wherein, the deep reconstruction network uses multi-scale feature extraction and global attention mechanism to suppress aerosol scattering noise.

[0010] In one embodiment of the present invention, the high-frequency pulse data packets are subjected to anti-jitter buffering and dynamic load reassembly to obtain target large packet data, which is then enqueued into a pre-set asynchronous circular queue, including: Allocate and lock a contiguous block of physical memory, and establish the asynchronous circular queue protected by a mutex lock on the physical memory; The high-frequency pulse data packets are captured by independent and parallel receiving threads, the payload of the data packets is extracted in real time, and the payload is spliced ​​into fixed lengths in the physical memory. When the accumulated load length reaches the preset output load threshold, the spliced ​​data is encapsulated into the target large packet data and written into the asynchronous circular queue.

[0011] In one embodiment of the present invention, transmitting the target large packet data in the asynchronous circular queue to the ground terminal according to a preset sending thread includes: The target large packet data is extracted from the asynchronous circular queue in a loop using independent and parallel sending threads; Based on the preset target throughput bit rate and target number of packets per second, combined with the actual byte length of the currently extracted target large packet data, the microsecond-level safe transmission waiting time interval required for the current transmission is dynamically calculated. During the microsecond-level safe transmission waiting time interval, a blocking sleep function is invoked to suspend the transmission thread.

[0012] In one embodiment of the present invention, based on a pre-set target throughput bit rate and target packets per second, and combined with the actual byte length of the currently extracted target large packet data, the microsecond-level secure transmission waiting time interval required for the current transmission is dynamically calculated, including: Based on the actual byte length of the target large packet data and the target throughput bit rate, calculate the first theoretical transmission time limited by bandwidth; Based on the target number of packets per second, calculate the second theoretical transmission time, which is limited by the packet rate; The maximum value between the first theoretical transmission time and the second theoretical transmission time is determined as the expected transmission time interval of the target large packet data; Get the system elapsed time from the last time the message was sent to the current time; Based on the difference between the expected transmission time interval and the system elapsed time, a microsecond-level safe transmission waiting time interval is determined.

[0013] In one embodiment of the present invention, the airborne terminal further includes a runtime state monitoring thread and an out-of-band serial port control thread independent of the sending thread and the receiving thread; the method further includes: Without system shutdown and without interruption of data flow, the out-of-band serial port control thread receives external configuration adjustment instructions and uses the system mutex lock to safely update the parameter variables in the runtime control structure, so that the sending thread applies the updated parameter variables in the next processing cycle. The parameter variables include the target throughput bit rate and / or the output load threshold.

[0014] In one embodiment of the present invention, for one or more received target large data packets, a three-dimensional photon counting matrix including temporal depth and spatial dimension is constructed, including: Extract photon count features from one or more target large packet data transmitted through the tunnel, and perform matrix mapping according to the physical correspondence between time sliding window, spatial height and spatial width; A three-dimensional photon counting matrix with dimensions B×C×D×H×W is constructed; where B is the batch size, C is the number of feature channels, D is the temporal depth dimension, and H and W are the height and width dimensions of the spatial image, respectively.

[0015] In one embodiment of the present invention, the deep reconstruction network includes an encoder path, a 3D global attention module, and a decoder path. The 3D photon counting matrix is ​​input into the pre-trained deep reconstruction network to obtain a 3D depth distance map of the detected target, including: Through the encoder path, the three-dimensional photon counting matrix is ​​controlled to flow through multiple spatial downsampling levels to extract local features, and a high-order spatiotemporal feature tensor is output at the deepest layer. The three-dimensional global attention module adds an attention mechanism to the high-order spatiotemporal feature tensor to obtain an enhanced feature sequence; wherein, the three-dimensional global attention module is used to suppress gamma distribution backscattering noise caused by haze aerosols. The enhanced feature sequence is reconstructed in reverse using the decoder path to obtain the fused features; The fused features are input into the depth map reconstruction branch to obtain a three-dimensional depth distance map of the detected target; wherein, the depth map reconstruction branch is used to independently perform probability normalization along the time depth axis, extract the temporal position of the Gaussian distribution target peak from the obtained single photon time distribution probability, and output a three-dimensional depth distance map.

[0016] In one embodiment of the present invention, an attention mechanism is added to the high-order spatiotemporal feature tensor through the three-dimensional global attention module to obtain an enhanced feature sequence, including: The input deepest high-order spatiotemporal feature tensor is divided into three-dimensional blocks and labeled, and then transformed into a one-dimensional feature vector sequence. Add three-dimensional positional encoding to the feature vector sequence to obtain the target feature vector sequence; The correlation weight matrix of the target feature vector sequence is calculated using a multi-head self-attention mechanism to suppress gamma-distributed backscattering noise caused by haze aerosols in a global view. Multiplying the correlation weight matrix by the feature vector sequence yields the enhanced feature sequence.

[0017] In one embodiment of the present invention, the enhanced feature sequence is reverse-reconstructed through the decoder path to obtain fused features, including: Following the reverse order of the three-dimensional block tokenization, the enhanced feature sequence is reverse-reshaped into a three-dimensional feature tensor with the same shape as the higher-order spatiotemporal feature tensor; By using 3D transposed convolution or 3D trilinear interpolation, the 3D feature tensor is upsampled step by step to gradually restore the resolution of the spatial dimension and the temporal depth dimension; After each upsampling stage, the upsampled features of the current level are spliced ​​and fused with the historical node features of the same resolution level and all shallower levels in the encoder path through a dense skip connection mechanism in the channel dimension until the original input resolution is restored, thus obtaining the fused features.

[0018] In one embodiment of the present invention, after the step of reversing the enhanced feature sequence through the decoder path to obtain the fused features, the method includes: The fused features are input into the confidence evaluation branch to obtain the confidence evaluation matrix of the detected target; wherein, the confidence evaluation branch is used to map the output of the confidence matrix consistent with the spatial resolution using the Sigmoid activation function, so as to quantify the physical reliability of the depth value of each pixel.

[0019] Secondly, the present invention provides a data processing system for single-photon fog-penetrating imaging, the system comprising an airborne terminal system and a ground terminal system; The airborne terminal system includes: The data acquisition module is used to acquire the raw echo signal of the single-photon detection system to the detection target, and to converge the raw echo signal to obtain multiple high-frequency pulse data packets; The first data processing module, connected to the data acquisition module, is used to perform anti-jitter buffering and dynamic load reassembly on the high-frequency pulse data packets to obtain the target large packet data and enqueue it into a pre-set asynchronous circular queue. An airborne wireless communication module, connected to the first data processing module, is used to transmit the target large packet data in the asynchronous circular queue to the ground terminal according to a preset sending thread; The ground terminal system includes a host computer, which includes: The second data processing module is used to construct a three-dimensional photon counting matrix, including temporal depth and spatial dimensions, for one or more target large data packets received. The deep reconstruction module, connected to the second data processing module, is used to input the three-dimensional photon counting matrix into a pre-trained deep reconstruction network to obtain a three-dimensional depth distance map of the detected target; wherein, the deep reconstruction network uses multi-scale feature extraction and global attention mechanism to suppress aerosol scattering noise.

[0020] In one embodiment of the present invention, the first data processing module includes: A network peripheral server, connected to the data acquisition module, is used to achieve transparent transmission and virtualization bridging of the underlying physical interface based on bus virtualization technology, and to access the high-frequency pulse data packets into the local area network. An airborne computer, whose input and output terminals are respectively connected to the network peripheral server and the airborne wireless communication module, is used to receive high-frequency pulse data packets passing through the local area network, and to perform anti-jitter buffering and dynamic load reassembly on the high-frequency pulse data at the airborne gateway layer to generate target large packet data, and send it into a pre-set asynchronous circular queue.

[0021] In one embodiment of the present invention, the ground terminal system further includes: The ground wireless communication module is used to establish a cross-terminal wireless data link with the airborne wireless communication module and to receive target large packet data sent by the airborne terminal system. The ground wireless communication module is connected to the host computer and is used to transmit the target large packet data to the host computer.

[0022] This invention provides a data processing method and system for single-photon fog-penetrating imaging. By introducing anti-jitter buffering and dynamic load reassembly mechanisms at the airborne end, it performs microsecond-level active flow shaping on sudden high-frequency pulse data spikes, transforming fragmented small packets into a smooth, constant data stream before transmission, thus physically avoiding overflow of the wireless bridge hardware buffer. Simultaneously, a three-dimensional photon counting matrix is ​​constructed at the ground end and input into a pre-trained deep reconstruction network. Multi-scale feature extraction and a three-dimensional global attention mechanism are used to forcibly suppress gamma-distributed backscattering noise caused by fog aerosols within a global field of view, ultimately synthesizing and outputting a high-precision three-dimensional depth-range map. Therefore, this invention solves the two major technical challenges of transmission link congestion and fog scattering noise without significantly increasing airborne hardware resource overhead, achieving highly reliable single-photon three-dimensional imaging in harsh environments with extremely low signal-to-noise ratios. Attached Figure Description

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a data processing method for single-photon fog-penetrating imaging according to an embodiment of the present invention.

[0025] Figure 2 This is a data processing system architecture diagram for single-photon fog-penetrating imaging provided in an embodiment of the present invention.

[0026] Figure 3 This is a flowchart illustrating the internal anti-jitter buffering and microsecond-level shaping processing of the airborne terminal, as provided in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the structure of a deep reconstruction network provided in an embodiment of the present invention.

[0028] Figure 5 This is a comparison chart of depth reconstruction effects under different fog concentrations provided in an embodiment of the present invention.

[0029] Figure 6 This is a schematic diagram of the structure of a data processing system for single-photon fog-penetrating imaging according to an embodiment of the present invention. Detailed Implementation

[0030] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0031] In a UAV-borne single-photon lidar detection system, the onboard single-photon detector is responsible for emitting laser pulses towards the target during flight and receiving photon echo signals reflected from the target surface and scattered by the environment. By accumulating and statistically analyzing multiple pulse echo events, a three-dimensional spatiotemporal photon count histogram can be constructed, thereby extracting the target's depth and distance information.

[0032] However, in actual hazy weather conditions, laser pulses undergo complex interactions with aerosol particles along their transmission path, resulting in strong backscattering. This backscattering manifests as a dominant gamma-ray scattering peak in the photon counting histogram, while the Gaussian echo peaks formed by reflections from the target surface are severely masked or even completely submerged. Furthermore, because single-photon detectors need to send massive amounts of high-frequency raw pulse data packets to the ground to support backend signal reconstruction, these data packets exhibit extremely strong burstiness and fragmentation characteristics.

[0033] In conventional wireless data pass-through schemes, airborne wireless bridges typically forward the raw pulse data packets generated by the detector directly. The internal hardware buffer resources of airborne wireless bridges are extremely limited, usually configured only with a small-capacity First-In-First-Out (FIFO) hardware queue. When a single-photon detector bursts with a large number of small UDP packets at an extremely high instantaneous rate, the packet arrival rate will far exceed the bridge's line-speed forwarding capability and buffer capacity limit within a very short time.

[0034] In this situation, the bridge hardware buffer will quickly become full and overflow, causing subsequent data packets to be discarded. Because the timing data of single-photon detection has a strict physical correspondence, once a photon counting pulse packet within a certain time window is lost during transmission, the photon distribution characteristics of that time window will undergo irreversible distortion—this distortion is not simply a decrease in signal-to-noise ratio, but rather the destruction of the physical structure of the photon arrival time series, causing subsequent deep reconstruction algorithms based on complete timing characteristics to lose their effective data foundation.

[0035] To address the aforementioned transmission congestion problem, an intuitive approach is to increase the hardware buffer capacity of the wireless transmission link or improve the transmission bandwidth. However, wireless communication modules suitable for UAV onboard platforms (such as the Silex 510 combined with the LQ3 wireless module) are limited by power consumption, size, and weight constraints, and their hardware buffer capacity and RF bandwidth have strict upper limits. Solutions that rely solely on hardware upgrades are difficult to implement in engineering on onboard platforms.

[0036] To address the aforementioned issue of scattering noise in haze, traditional peak detection algorithms can only perform local peak searches along the time axis of the photon count histogram for a single pixel. However, under conditions of extremely low signal-to-noise ratio, such as dense fog, the gamma-distributed scattering peaks generated by aerosol backscattering are much larger in amplitude than the target Gaussian echo peaks. Local peak detection algorithms are prone to misidentifying scattering peaks as target echoes, resulting in a large number of scattering noise points filling the depth map. Even when using conventional image dehazing or spatial filtering algorithms, their processing scope is usually limited to a pixel window within the spatial neighborhood, lacking the ability to utilize the correlation information of photon events between nodes across time and space in the entire image. This makes it difficult to robustly recover the spatial contour and depth information of the target from 3D photon count data heavily contaminated by scattering noise.

[0037] Based on this, this application provides a data processing method and system for single-photon fog-penetrating imaging. By introducing anti-jitter buffering and dynamic load reassembly mechanisms at the airborne end, microsecond-level active flow shaping is performed on sudden high-frequency pulse data surges, transforming fragmented small packets into a smooth and constant data stream before transmission, thus physically avoiding overflow of the wireless bridge hardware buffer. Simultaneously, a three-dimensional photon counting matrix is ​​constructed at the ground end and input into a pre-trained deep reconstruction network. Multi-scale feature extraction and a three-dimensional global attention mechanism are used to forcibly suppress gamma-distributed backscattering noise caused by fog aerosols under a global field of view, ultimately synthesizing and outputting a high-precision three-dimensional depth-range map. Thus, this invention solves the two major technical challenges of transmission link congestion and fog scattering noise without significantly increasing airborne hardware resource overhead, achieving highly reliable single-photon three-dimensional imaging in harsh environments with extremely low signal-to-noise ratios.

[0038] Please see Figure 1 This invention provides a data processing method for single-photon fog-penetrating imaging. This method is used to perform end-to-end processing of signals acquired by a single-photon detection system mounted on a UAV in complex weather environments such as fog and haze, to achieve high-precision three-dimensional imaging. Specifically, this data processing method is applied to a data processing system including an airborne end and a ground end. The airborne end is deployed in the UAV payload bay and is used to perform signal acquisition, buffering and shaping, and data transmission; the ground end is deployed in a ground data processing center and is used to perform data reconstruction and depth reconstruction.

[0039] The method may include the following steps.

[0040] Step S110: The airborne terminal acquires the original echo signal of the single-photon detection system for the target and converges the original echo signal to obtain multiple high-frequency pulse data packets.

[0041] In this embodiment, the airborne unit is equipped with a single-photon detection system. This system emits laser pulses toward the target and receives photon echo signals reflected from the target surface and scattered by the environment. Specifically, the single-photon detection system can use a Geiger-mode avalanche photodiode (GM-APD) as the core detection device, which has extremely high detection sensitivity at the single-photon level.

[0042] The raw echo signal is the original electrical signal collected by the single-photon detection system during the detection process, containing information about the photon arrival time. Specifically, the raw echo signal contains both the target echo component formed by reflection from the surface of the target and the scattering noise component formed by backscattering from haze aerosol particles.

[0043] High-frequency pulse data packets are data units formed after the original echo signal has been digitized and encapsulated. Specifically, a single-photon detection system can encapsulate information such as photon arrival timestamps and pulse sequence numbers from the original echo signal into User Datagram Protocol (UDP) data packets. Because the single-photon detection system emits laser pulses and receives echoes at extremely high repetition rates, the resulting high-frequency pulse data packets possess extremely strong burst and high-frequency characteristics, and the data packet arrival rate can reach its peak within a short period of time.

[0044] The airborne terminal uses a bus expansion module or high-speed serial interface to aggregate multiple pulse signals or low-level control signals generated by the single-photon detection system to form a unified data stream for subsequent modules to process. In some embodiments, the airborne terminal can use a high-speed time-to-digital converter to measure the time of multi-channel echo signals and aggregate the measurement results into a high-frequency pulse data packet.

[0045] Step S120: Perform anti-jitter buffering and dynamic load reassembly on the high-frequency pulse data packet to obtain the target large packet data, and enqueue it into a pre-set asynchronous circular queue.

[0046] In this embodiment, the jitter buffer is a mechanism used to eliminate time jitter introduced by operating system scheduling or memory management during data transmission. Specifically, the jitter buffer can be implemented by locking contiguous memory regions at the physical memory level and shielding the operating system's virtual memory page swapping scheduling, thereby ensuring the time determinism of data during the buffering process.

[0047] Dynamic load reassembly is a processing mechanism used to reassemble fragmented small data packets into larger data packets adapted to downstream transmission links. Specifically, dynamic load reassembly may involve real-time extraction of the payload from high-frequency pulse data packets, fixed-length splicing of the payload in memory, and encapsulation of the spliced ​​data into the target large packet when the accumulated payload length reaches a preset output load threshold. Through dynamic load reassembly, network protocol stack header redundancy caused by high-frequency small packets can be eliminated, reducing the impact of data packet rate on downstream hardware buffers.

[0048] The target large packet data is a data unit generated by extracting the payload from high-frequency pulse data packets and splicing them to a fixed length, and is adapted to the transmission characteristics of downstream wireless links. Specifically, the payload length of the target large packet data can be set according to the throughput characteristics of the downstream wireless bridge module and the wireless transmitting terminal, so that the shaped data stream strictly matches the transmission capacity of the downstream physical link.

[0049] An asynchronous circular queue is a queue structure used to implement thread-safe data buffering internally on the device. Specifically, an asynchronous circular queue is protected by a mutex lock, allowing independent receiving and sending threads to perform data writing and reading operations in parallel without data contention. Asynchronous circular queues are pre-allocated and locked at the physical memory level, avoiding the system overhead and latency caused by dynamic memory allocation at runtime.

[0050] In some embodiments, the onboard unit can allocate a contiguous anonymous private memory region during the initialization phase via a kernel-level system call, and call the underlying memory locking interface to force this contiguous memory region to reside in physical memory, thereby establishing an asynchronous circular queue. After capturing high-frequency pulse data packets, the receiving thread performs fixed-length splicing. When the accumulated payload length reaches the output load threshold, it encapsulates and generates the target large packet data, and safely enqueues it into the asynchronous circular queue.

[0051] Step S130: Transmit the target large packet data in the asynchronous circular queue to the ground terminal according to the preset sending thread.

[0052] In this embodiment, the preset sending thread is a parallel processing thread that runs independently of the receiving thread. It is used to cyclically extract target large packet data from the asynchronous circular queue and control the data transmission rate according to microsecond-level time intervals. Specifically, the preset sending thread can dynamically calculate the required microsecond-level safe transmission waiting time interval based on the preset target throughput bit rate and the number of target packets per second, combined with the actual byte length of the currently extracted target large packet data. Within this time interval, it calls a microsecond-level blocking sleep function to precisely suspend the data transmission, thereby transforming the sudden high-frequency pulse data surge at the front end into a smooth and constant output data stream.

[0053] The transmission to the ground can be achieved through a multi-level heterogeneous physical link on the airborne end. Specifically, the multi-level heterogeneous physical link can include a multi-level physical media bridge from the underlying serial bus and local area network Ethernet to space radio frequency microwave. The shaped constant data stream is transmitted to the wireless bridge module via the airborne local area network interface, where a Media Access Control Layer / Internet Protocol Layer (MAC / IP Layer) protocol conversion is performed to adapt to the airborne physical interface. It is then transparently forwarded to the backend wireless transmitting terminal and finally transmitted to the ground end via air interface electromagnetic waves. Relying on the aforementioned microsecond-level active shaping mechanism, this constant data stream strictly matches the throughput bottlenecks of the wireless bridge module and the wireless transmitting terminal, avoiding hardware buffer overflow at the physical level.

[0054] Step S210: Using the ground terminal, construct a three-dimensional photon counting matrix including temporal depth and spatial dimensions for one or more target large packet data received.

[0055] In this embodiment, the ground terminal receives one or more target data packets transmitted without loss and extracts photon counting features from the target data packets. Specifically, the photon counting features are used to characterize the cumulative number of photon events detected at a specific spatial location within a specific time window.

[0056] A three-dimensional photon counting matrix is ​​a multi-dimensional data structure formed by mapping photon counting characteristics according to temporal and spatial dimensions. Specifically, the three-dimensional photon counting matrix can be mapped according to the physical correspondence between time sliding windows, spatial height, and spatial width. The temporal depth dimension corresponds to the flight time of the laser pulse from emission to reception, reflecting the distance information of each point on the surface of the target; the spatial height and spatial width dimensions correspond to the spatial scanning range of the single-photon detection system, reflecting the two-dimensional spatial distribution of the target.

[0057] In some embodiments, the dimensions of the three-dimensional photon counting matrix can be constructed as B×C×D×H×W. Here, B is the batch size, C is the number of feature channels, D is the temporal depth dimension, and H and W are the height and width dimensions of the spatial image, respectively. By constructing the three-dimensional photon counting matrix, the originally linearly transmitted data stream is restored to a structured data representation corresponding to the detection physical process, providing a standardized input format for subsequent deep reconstruction networks.

[0058] Step S220: Input the three-dimensional photon counting matrix into the pre-trained deep reconstruction network to obtain the three-dimensional depth distance map of the detected target; wherein, the deep reconstruction network uses multi-scale feature extraction and global attention mechanism to suppress aerosol scattering noise.

[0059] Deep reconstruction networks are deep learning models used to recover the 3D depth information of a target from 3D photon count data heavily contaminated by aerosol scattering noise from smog. Specifically, deep reconstruction networks utilize multi-scale feature extraction and a global attention mechanism to suppress aerosol scattering noise.

[0060] Multi-scale feature extraction is a processing mechanism used to capture target features in photon counting data at different spatial and temporal resolutions. Specifically, multi-scale feature extraction can be achieved through multiple spatial downsampling levels in the encoder path. After each level performs local feature extraction on the input features, the spatial and temporal resolution is progressively reduced through pooling operations, thereby implicitly extracting multi-scale local spatial prior information of the target and its surrounding environment, and outputting a high-order spatiotemporal feature tensor at the deepest layer.

[0061] The global attention mechanism is a processing mechanism used to calculate global association weights between nodes across the entire spatiotemporal region of 3D photon counting data. Specifically, the global attention mechanism can be achieved by dividing and serializing the deepest high-order spatiotemporal feature tensor into blocks, injecting it into 3D position encoding, and then feeding it into a cascaded multi-head self-attention module. Scaling dot product attention is then used to calculate the global association weights between nodes across the entire spatiotemporal region of the image. Since backscattering noise caused by haze aerosols exhibits specific gamma distribution statistical characteristics in the 3D photon counting matrix, while the target reflection signal exhibits Gaussian distribution statistical characteristics, the global attention mechanism can learn and identify the differences between the two distribution patterns from a global perspective. This allows it to forcibly suppress the gamma-distributed backscattering noise caused by haze aerosols and enhance the characteristic representation of the target Gaussian distribution echo signal.

[0062] A 3D depth distance map is a three-dimensional image output by a depth reconstruction network, containing distance information between various spatial points on the target surface and the single-photon detection system. Specifically, the 3D depth distance map is obtained by independently performing probability normalization along the time-depth axis through the depth map reconstruction branch, transforming the temporal network response into a single-photon temporal distribution probability, and extracting the temporal position of the Gaussian distribution target peak from it. The temporal position is proportional to the flight time of the laser pulse, thereby allowing the calculation of the physical distance corresponding to each spatial point.

[0063] In some embodiments, the deep reconstruction network can adopt an encoder-decoder structure based on a densely nested architecture, introduce a three-dimensional global attention module in the deepest layer, and use a dense skip connection mechanism in the decoding stage to achieve efficient fusion of deep global features and shallow local features. Finally, the three-dimensional depth distance map and confidence evaluation matrix are output through dual-branch output heads.

[0064] The data processing method provided in this application transforms sudden high-frequency pulse data spikes into smooth, constant data streams through anti-jitter buffering and dynamic load reassembly at the airborne end, physically avoiding the hardware buffer overflow problem of the wireless transmission link. At the ground end, by constructing a three-dimensional photon counting matrix from the lossless data and inputting it into a depth reconstruction network that uses multi-scale feature extraction and global attention mechanisms to suppress aerosol scattering noise, high-precision three-dimensional depth range map output is achieved in harsh environments with extremely low signal-to-noise ratios. Thus, this application embodiment solves the two major technical challenges of transmission link congestion and fog scattering noise without significantly increasing airborne hardware resource overhead, achieving end-to-end high-reliability single-photon fog-penetrating imaging.

[0065] In some embodiments, step S120, which involves performing anti-jitter buffering and dynamic load reassembly on the high-frequency pulse data packets to obtain target large packet data, and queuing it into a pre-set asynchronous circular queue, may include the following steps.

[0066] Step S121: Allocate and lock a contiguous block of physical memory, and establish the asynchronous circular queue protected by a mutex lock on the physical memory.

[0067] Please see Figure 2 and Figure 3 In this embodiment, during the initialization phase, the airborne transmission gateway allocates a contiguous anonymous private memory region through a kernel-level system call and calls the underlying memory locking interface to force the contiguous memory region to reside in physical memory, thereby shielding the operating system's virtual memory page swapping scheduling mechanism from the hardware level and establishing the asynchronous circular queue.

[0068] Physical memory is the actual random access memory region in an onboard computer that is directly connected to the processor bus. Specifically, by forcing contiguous memory regions to reside in physical memory, the operating system can avoid swapping the contents of this memory region to external storage media such as disks during operation, thereby ensuring that data is not affected by the latency jitter of virtual memory page scheduling during the buffering process.

[0069] A mutex lock is a synchronization control mechanism used to implement concurrent access to shared resources by multiple threads. Specifically, a mutex lock ensures that only one thread can perform write or read operations on an asynchronous circular queue at any given time, thereby preventing data races and inconsistencies caused by multiple parallel threads operating on the queue simultaneously.

[0070] The asynchronous circular queue is a thread-safe buffer queue with circular read and write pointers, hosted in the locked physical memory. Specifically, the asynchronous circular queue internally maintains a read pointer and a write pointer. The receiving thread writes data from the write pointer position and advances the write pointer, while the sending thread reads data from the read pointer position and advances the read pointer. When either pointer reaches the end of the queue, it automatically wraps back to the beginning of the queue, thus achieving continuous data buffering within a fixed-size memory space. The queue depth and the memory block size of each queue element can be determined based on the total locked memory capacity and the preset wireless output load size.

[0071] Step S122: The high-frequency pulse data packets are captured by independent parallel receiving threads, the payload of the data packets is extracted in real time, and the payload is spliced ​​into fixed lengths in the physical memory.

[0072] In this embodiment, please continue to refer to Figure 3 The airborne transmission gateway allocates an independent, parallel receiving thread. This receiving thread calls the bulk network packet receiving interface to capture high-frequency raw User Datagram Protocol (UDP) pulse packets generated by the single-photon detector and pushes them into the reassembly module.

[0073] The payload is the actual data content of a high-frequency pulse data packet, excluding the network protocol stack header, and includes physical information such as the photon arrival timestamp and pulse sequence number. Specifically, after capturing a high-frequency pulse data packet, the receiving thread extracts the payload portion of the packet in real time and discards the network protocol stack header to reduce redundant overhead in subsequent processing and transmission.

[0074] Fixed-length splicing is a process of splicing continuously received payloads into a fixed length in physical memory. Specifically, the reassembly module appends each extracted payload sequentially to a pre-allocated splicing buffer in physical memory and records the current cumulative payload length in real time.

[0075] Step S123: When the accumulated load length reaches the preset output load threshold, the spliced ​​data is encapsulated into the target large packet data and written into the asynchronous circular queue.

[0076] In this embodiment, the output load threshold is a target data packet payload length that is pre-set based on the transmission characteristics of the downstream wireless link and adapted to the throughput capabilities of the wireless bridge module and the wireless transmitting terminal. Specifically, the output load threshold can be determined based on one or more factors such as the transmission bandwidth of the wireless link, the data packet processing rate, and the bit error rate characteristics, so as to achieve a balance between transmission efficiency and transmission reliability for the shaped target large packet data.

[0077] When the length of the accumulated load reaches the output load threshold, the reassembly module encapsulates the data in the splicing buffer into the target large packet data and securely enqueues it into the pre-divided anonymous private memory block in the asynchronous circular queue. This eliminates the redundancy of the network protocol stack header caused by high-frequency small packets and avoids the system overhead and latency caused by dynamic memory allocation.

[0078] Through the aforementioned anti-jitter buffering and dynamic load reassembly mechanisms, on the one hand, physical memory locking shields the time jitter introduced by the operating system's virtual memory page swapping at the hardware level, ensuring the time determinism of the buffering process; on the other hand, the small packet splicing and large packet reassembly strategy transforms the originally fragmented high-frequency pulse data packets into whole large-size data packets, significantly reducing the number of data packets that need to be sent per unit time, thereby alleviating the pressure on the downstream wireless bridge hardware buffer and laying the data format foundation for the subsequent implementation of active traffic shaping.

[0079] In some embodiments, step S130, transmitting the target large packet data in the asynchronous circular queue to the ground terminal according to a preset sending thread, may include the following steps.

[0080] Step S131: Extract the target large packet data from the asynchronous circular queue in a loop using independent and parallel sending threads.

[0081] Please continue reading. Figure 3 In this embodiment, the airborne transmission gateway allocates an independent and parallel sending thread. This sending thread runs independently of the aforementioned receiving thread, and the two perform thread-safe data handover through the asynchronous circular queue. The sending thread cyclically extracts the encapsulated target packet data from the asynchronous circular queue and sends it to the rate limiting module.

[0082] Step S132: Based on the preset target throughput bit rate and target number of packets per second, and combined with the actual byte length of the currently extracted target large packet data, dynamically calculate the microsecond-level secure transmission waiting time interval required for the current transmission.

[0083] In this embodiment, the target throughput bit rate is a control parameter used to limit the total number of bits of data transmitted to the ground end per unit time. Specifically, the target throughput bit rate can be set according to the actual transmission bandwidth of the downstream wireless bridge module and the wireless transmitting terminal, so as to ensure that the shaped data stream rate does not exceed the maximum throughput capacity of the wireless link, thereby avoiding congestion in the wireless transmission process.

[0084] Packets per second (TPS) is a control parameter used to limit the number of data packets sent to the ground station per unit of time. Specifically, TPS can be set according to the hardware buffer capacity and packet processing rate of the downstream wireless bridge module. Since each arriving packet occupies a queue entry in the bridge's hardware buffer, limiting the number of packets sent per second ensures that the bridge's hardware buffer always has enough free entries to accommodate newly arriving packets, thus mitigating the risk of buffer overflow at the source.

[0085] The microsecond-level secure transmission wait interval is the length of time that the sending thread must actively wait after the current target large packet data transmission is completed and before the next target large packet data transmission begins. Specifically, the microsecond-level secure transmission wait interval is dynamically calculated by the rate limiting module based on the target throughput bit rate, the number of target packets per second, and the actual byte length of the currently extracted target large packet data.

[0086] In some embodiments, the rate limiting module first calculates a first theoretical transmission time limited by bandwidth and a second theoretical transmission time limited by packet rate, based on the actual byte length of the target large packet data. Then, the maximum value of the two is taken as the expected transmission time interval for the current target large packet data. Finally, the actual system elapsed time from the completion of the last transmission to the current moment is obtained, and a microsecond-level safe transmission waiting time interval is determined based on the difference between the expected transmission time interval and the system elapsed time.

[0087] Step S133: During the microsecond-level safe transmission waiting time interval, call the blocking sleep function to suspend the transmission thread.

[0088] In this embodiment, the blocking sleep function is a high-precision delay function provided by the operating system, used to suspend the currently calling thread for a specified time. Specifically, the blocking sleep function can be a nanosecond-level sleep function. Its underlying working principle is as follows: when the sending thread calls the blocking sleep function, the sending thread actively relinquishes its time slice execution right in the central processing unit, avoiding the use of a busy-wait polling mode that consumes a lot of system resources. At this time, the operating system kernel switches the state of the sending thread from the running state to the blocked suspended state and registers a high-precision timer with the system kernel that matches the microsecond-level safe transmission waiting time interval. When the underlying hardware clock of the system reaches the preset time interval and triggers an interrupt, the operating system scheduler will capture the interrupt signal and wake up the sending thread back to the ready queue, thereby achieving precise delayed transmission that strictly follows the microsecond-level time interval with extremely low central processing unit usage.

[0089] Through the aforementioned microsecond-level active shaping and smoothing transmission mechanism, the transmission thread delivers the shaped target large packet data to the downstream airborne bridge and wireless communication module in a packet-by-packet, continuous, and uniform manner. Utilizing an asynchronous circular queue as a deep software-level buffer, it absorbs and temporarily stores instantaneous high-frequency pulse signals from the front-end single-photon detection system, transforming sudden high-frequency pulse data spikes from the front end into a smooth, uniform, and constant output data stream. This constant data stream precisely matches the throughput bottlenecks of the wireless bridge module and the wireless transmitting terminal, thereby physically avoiding hardware buffer overflow and ensuring lossless data transmission to the ground terminal system.

[0090] In some embodiments, step S132, based on a preset target throughput bit rate and target packets per second, and combined with the actual byte length of the currently extracted target large packet data, dynamically calculates the microsecond-level secure transmission waiting time interval required for the current transmission, which may include the following steps.

[0091] Step S1321: Calculate the first theoretical transmission time limited by bandwidth based on the actual byte length of the target large packet data and the target throughput bit rate.

[0092] In this embodiment, the first theoretical transmission time is the theoretical time required to send the current target large packet data, considering only the wireless link bandwidth constraint. Specifically, the first theoretical transmission time can be calculated based on the actual byte length of the target large packet data and the target throughput bit rate. In a specific example, the first theoretical transmission time... The calculation formula is: (microseconds). Among them, The actual byte length of the target large packet data. The target throughput bit rate is used, and the calculation results are in microseconds. The first theoretical transmission time reflects the minimum time window required to send the target large packet of data in terms of wireless link bandwidth.

[0093] Step S1322: Calculate the second theoretical transmission time, which is limited by the packet rate, based on the target number of packets per second.

[0094] In this embodiment, the second theoretical transmission time is the theoretical time required to send the current target large packet data, considering only the packet processing rate constraint of the wireless bridge hardware buffer. Specifically, the second theoretical transmission time can be calculated based on the number of target packets per second. In a specific example, the second theoretical transmission time... The calculation formula is: (microseconds). Here, PPS represents the target packets per second, and the calculation result is in microseconds. The second theoretical transmission time reflects the minimum transmission interval required between two adjacent data packets in terms of packet processing rate.

[0095] Step S1323: Determine the maximum value between the first theoretical transmission time and the second theoretical transmission time as the expected transmission time interval of the target large packet data.

[0096] In this embodiment, the expected transmission time interval is the ideal time interval required to transmit the current target large packet data, determined after comprehensively considering bandwidth constraints and packet rate constraints. Specifically, the rate limiting module determines the maximum value between the first theoretical transmission time and the second theoretical transmission time as the expected transmission time interval, i.e. By taking the maximum of the two, it can be ensured that the transmission interval of the current target large packet data simultaneously meets the wireless link bandwidth constraint and the bridge hardware buffer packet processing rate constraint, avoiding exceeding the capacity of the downstream physical link in either dimension.

[0097] Step S1324: Obtain the system elapsed time from the last transmission to the current time.

[0098] In this embodiment, the system elapsed time is the actual elapsed time between the last completion time of transmission and the current calculation time. Specifically, the rate limiting module can obtain the actual time difference between the timestamp of the last transmission thread completing the transmission of the target large packet data and the current time by calling the system's high-precision timer interface. This time difference is measured in microseconds and denoted as . .

[0099] Step S1325: Determine the microsecond-level safe transmission waiting time interval based on the difference between the expected transmission time interval and the system elapsed time.

[0100] In this embodiment, the microsecond-level safe transmission waiting time interval is the remaining time that the sending thread needs to actively wait before starting the next transmission operation. Specifically, the rate limiting module determines the microsecond-level safe transmission waiting time interval based on the difference between the expected transmission time interval and the system elapsed time. When the system elapsed time exceeds the expected transmission interval, it indicates that the actual processing time of the current target large packet data has met the transmission interval requirement. At this point, the microsecond-level safe transmission waiting interval is zero, and the sending thread can immediately send the next target large packet data without waiting. When the system elapsed time has not yet reached the expected transmission interval, the sending thread needs to call the blocking sleep function within the remaining microsecond-level safe transmission waiting interval to precisely suspend the transmission, in order to make up the actual transmission interval to the expected transmission interval.

[0101] Through the aforementioned dynamic calculation mechanism based on two-layer constraints, the rate limiting module can adaptively adjust the sending interval according to the actual byte length of each target large packet of data, so that the final output data stream exhibits a uniform and smooth rate characteristic in the time dimension, thereby maximizing data throughput efficiency while adapting to the transmission capacity of downstream physical links.

[0102] In some embodiments, the airborne terminal further includes a runtime state monitoring thread and an out-of-band serial port control thread, independent of the sending thread and the receiving thread. The data processing method may also include the following steps.

[0103] Step S134: Without system shutdown and data flow interruption, the external configuration adjustment command is received through the out-of-band serial port control thread, and the parameter variables in the runtime control structure are safely updated using the system mutex lock, so that the sending thread applies the updated parameter variables in the next processing cycle. The parameter variables include the target throughput bit rate and / or the output load threshold.

[0104] In this embodiment, please continue to refer to Figure 3 In this embodiment, in addition to the aforementioned receiving thread and sending thread, the airborne transmission gateway also has an independent runtime status monitoring thread and an out-of-band serial port control thread.

[0105] The runtime status monitoring thread is an independent thread used to collect and report the operational status of the airborne transmission gateway in real time. Specifically, the runtime status monitoring thread can periodically collect one or more runtime status indicators such as the current data throughput rate, the queue occupancy depth of the asynchronous circular queue, the blocking and waiting statistics of the sending thread, and the CPU utilization rate, and report them to the external control terminal through the out-of-band communication interface, providing operators with visual feedback on the system's operational status.

[0106] The out-of-band serial port control thread is independent of the data transmission channel and is dedicated to receiving external configuration adjustment commands. Specifically, the out-of-band serial port control thread is isolated from the data transmission channel in both physical interface and logical path. Its process of receiving external configuration adjustment commands does not consume bandwidth resources of the data transmission link, nor does it compete with the real-time data transmission and reception operations of the receiving and sending threads.

[0107] External configuration adjustment commands are control commands used to dynamically modify transmission shaping parameters during system operation. Specifically, external configuration adjustment commands may include one or more commands such as adjusting the transmission interval, adjusting the output load, adjusting the statistical output frequency, and logging on / off commands. Operators can send corresponding configuration adjustment commands through an external control terminal based on real-time changes in the UAV flight environment (such as changes in channel quality and transmission distance) to dynamically optimize the underlying transmission shaping mechanism in real time.

[0108] The runtime control structure is a data structure in the memory of the airborne transmission gateway used to centrally store transmission shaping-related parameter variables. Specifically, the runtime control structure contains parameter variables such as target throughput bit rate and output load threshold. When executing their respective data processing logic, the reassembly module of the receiving thread and the rate limiting module of the sending thread both read the current parameter values ​​from this runtime control structure.

[0109] In some embodiments, when the out-of-band serial port control thread receives an external configuration adjustment command, it safely updates the parameter variables in the runtime control structure using a system mutex lock, without system downtime or interruption of data flow. Specifically, before updating the parameter variables, the out-of-band serial port control thread acquires the mutex lock associated with the runtime control structure. After successful acquisition, it performs a write operation on the parameter variables and immediately releases the mutex lock after the write is complete. Through the synchronization protection of the mutex lock, it is ensured that the parameter variable update operation will not compete with the receiving thread or the sending thread's read operation on the runtime control structure, thereby guaranteeing the atomicity and consistency of the parameter update.

[0110] The next processing cycle refers to the time point at which the receiving thread or sending thread begins executing the next round of data processing after the current data processing task is completed. Specifically, when the parameter variables in the runtime control structure are updated, the rate limiting module immediately reads and applies the updated parameter variables (such as the new target throughput bit rate) in the next execution cycle of its sending thread; the reassembly module immediately reads and applies the updated parameter variables (such as the new output load threshold) in the next execution cycle of its receiving thread. During this process, the system reuses the allocated global asynchronous circular queue without reallocating memory resources, thus achieving a smooth transition for hot parameter switching.

[0111] Through the aforementioned dynamic runtime parameter tuning mechanism based on an independent control link, operators can adjust core transmission shaping parameters such as target throughput bit rate and / or output load threshold online according to the real-time changing flight environment and channel conditions, without system shutdown or data interruption. This enables the airborne transmission gateway to flexibly adapt to complex and time-varying UAV flight operation scenarios, maximizing transmission efficiency while ensuring lossless data transmission.

[0112] In some embodiments, in step S210, constructing a three-dimensional photon counting matrix including temporal depth and spatial dimension for one or more received target large packet data may include the following steps.

[0113] Step S211: Extract the photon count features from the data of one or more target packets that have been transmitted through the transmission, and perform matrix mapping according to the physical correspondence between time sliding window, spatial height and spatial width.

[0114] Please see Figure 2 In this embodiment, the ground terminal receives one or more target data packets transmitted losslessly from the airborne terminal system via a ground wireless communication module and transmits them to the host computer. The host computer parses the received target data packets to construct a three-dimensional photon counting matrix.

[0115] Photon counting characteristics are statistical features representing the cumulative number of single-photon events detected at a specific spatial location within a specific time window. Specifically, since a single-photon detection system emits laser pulses at a high repetition rate towards the target, each pixel location corresponds to a pulse sequence, and the return signal of each pulse is recorded by the single-photon detector as a photon arrival event. By accumulating and statistically analyzing the results of multiple repeated pulse measurements, the cumulative photon count value at different temporal depths for each pixel location can be obtained; this cumulative value is the photon counting characteristic. In a clear, fog-free environment, the photon counting characteristic of the target surface reflection signal exhibits a Gaussian distribution along the time-depth axis. In environments with high aerosol concentrations, such as fog or haze, due to the influence of backscattering, the photon counting characteristic will also be superimposed with a scattering noise component exhibiting a gamma distribution.

[0116] A time-sliding window is a method of discretizing continuous photon arrival time information into multiple time intervals. Specifically, the width of the time-sliding window is matched to the time resolution of the time-to-digital converter of the single-photon detection system. Each time-sliding window corresponds to a specific photon flight time interval, that is, a specific distance interval between the target surface and the single-photon detection system. The time-sliding windows are arranged continuously on the time axis, covering the complete time range from the laser pulse emission time to the echo reception time corresponding to the maximum detection distance.

[0117] Spatial height and spatial width are two-dimensional spatial dimensions used to describe the spatial scanning range of a single-photon detection system. Specifically, spatial height corresponds to the number of scan rows in the scanning direction of the single-photon detection system, and spatial width corresponds to the number of pixels in each scan row. The product of spatial height and spatial width determines the total number of pixels contained in a frame of 3D point cloud data, with each pixel corresponding to a specific spatial location on the surface of the target being detected.

[0118] Matrix mapping is the process of organizing photon counting features into a multidimensional matrix according to physical correspondences. Specifically, the second data processing module extracts the photon counting features contained in each target large packet of data, and fills the photon counting features into the corresponding positions in the three-dimensional matrix according to the time window index to which the photon arrival time belongs and the spatial scan position corresponding to the photon event. Here, the time window index corresponds to the time depth dimension, the spatial scan row index corresponds to the spatial height dimension, and the spatial scan column index corresponds to the spatial width dimension. Through matrix mapping, the photon counting information, which originally existed as a linear data stream during transmission, is restored into a structured multidimensional data representation corresponding to the detection physical process.

[0119] Step S212: Construct a three-dimensional photon counting matrix with dimensions B×C×D×H×W; where B is the batch size, C is the number of feature channels, D is the temporal depth dimension, and H and W are the height and width dimensions of the spatial image, respectively.

[0120] In some embodiments, the dimensions of the three-dimensional photon counting matrix are constructed as a five-dimensional tensor of B×C×D×H×W. Here, B is the batch size, representing the number of samples input into the three-dimensional photon counting matrix of the depth reconstruction network in a single operation. The batch size setting balances the inference efficiency of the depth reconstruction network with the constraints of the graphics processor's memory capacity. C is the number of feature channels, representing the number of feature dimensions at each pixel location at each temporal depth. In single-photon detection scenarios, the number of feature channels is typically set to 1, meaning it only includes the cumulative photon count. D is the temporal depth dimension, corresponding to the total number of time windows, i.e., the total time step. The value of D determines the depth range and distance resolution that the depth reconstruction network can resolve. H is the height dimension of the spatial image, corresponding to the number of scan rows in the single-photon detection system. W is the width dimension of the spatial image, corresponding to the number of scan pixels per row in the single-photon detection system.

[0121] By constructing the aforementioned three-dimensional photon counting matrix with dimensions B×C×D×H×W, the linear data stream received from the ground end is transformed into a structured tensor representation corresponding to the physical process of single-photon detection. This provides a standardized input format for the subsequent deep reconstruction network, enabling the network to simultaneously utilize photon distribution statistics in both the temporal and spatial dimensions for feature extraction and noise suppression.

[0122] In some embodiments, please refer to Figure 4 The deep reconstruction network includes an encoder path, a 3D global attention module, and a decoder path. The deep reconstruction network employs an encoder-decoder structure based on a densely nested architecture, and introduces a 3D global attention module between the deepest layer of the encoder path and the deepest layer of the decoder path to form a complete feature extraction, global association, and deep reconstruction processing chain.

[0123] In step S220, the three-dimensional photon counting matrix is ​​input into a pre-trained depth reconstruction network to obtain a three-dimensional depth distance map of the detected target, which may include the following steps.

[0124] Step S221: Through the encoder path, control the three-dimensional photon counting matrix to flow through multiple spatial downsampling levels to extract local features, and output a high-order spatiotemporal feature tensor at the deepest layer.

[0125] In this embodiment, the encoder path is a network structure in the deep reconstruction network used to extract multi-scale local spatiotemporal features from the three-dimensional photon counting matrix level by level. Specifically, after the three-dimensional photon counting matrix is ​​input into the encoder path, it flows through multiple spatial downsampling levels sequentially. For example... Figure 4 As shown in part (b), each spatial downsampling level uses a 3D residual convolutional block containing a 3D convolutional layer, a 3D batch normalization layer, and a linear rectified activation function for local feature extraction, and then progressively reduces the spatial and temporal depth resolution through a 3D max pooling layer. The mathematical formula for the 3D residual convolutional block is expressed as follows:

[0126] In the formula, The feature tensor input to the current residual convolutional block, To output the feature tensor, This represents a nonlinear transformation operator composed of a concatenation of three-dimensional convolution and three-dimensional batch normalization. It is a linear rectified activation function. The residual connection introduced by the addition operator effectively avoids the degradation of spatiotemporal features during forward propagation.

[0127] The higher-order spatiotemporal feature tensor is a multidimensional feature representation containing highly abstract semantic information, output from the deepest layer of the encoder path. Specifically, the higher-order spatiotemporal feature tensor is a feature tensor obtained after progressive feature extraction through multiple spatial downsampling levels. It is compressed to a lower resolution in both spatial and temporal depth dimensions, but the feature channel dimension is expanded to a higher dimension, thus implicitly containing local spatial prior information about the target and its surrounding environment at multiple scales in a compact form.

[0128] Step S222: Through the three-dimensional global attention module, an attention mechanism is added to the high-order spatiotemporal feature tensor to obtain an enhanced feature sequence; wherein, the three-dimensional global attention module is used to suppress gamma distribution backscattering noise caused by haze aerosols.

[0129] In this embodiment, as Figure 4 As shown in section (c), the 3D global attention module is a network structure in the deep reconstruction network used to calculate the association weights between nodes across the entire spatiotemporal region of the map under a global perspective. The 3D global attention module first performs dynamic feature block labeling on the deepest high-order spatiotemporal feature tensor, transforming it into a one-dimensional spatial-temporal feature vector sequence, and injects 3D positional encoding into this feature vector sequence. Subsequently, under the stabilization of root mean square normalization, a multi-head self-attention mechanism is used to calculate the global association weights between nodes across the entire spatiotemporal region of the map. Since the backscattering noise caused by haze aerosols exhibits specific gamma distribution statistical characteristics in the 3D photon counting matrix, while the target reflection signal exhibits Gaussian distribution statistical characteristics, the multi-head self-attention mechanism can learn and identify the differences between the two distribution patterns across the entire map from a global perspective, thereby forcibly suppressing the gamma-distributed backscattering noise caused by haze aerosols.

[0130] After self-attention computation, the feature vector sequence is non-linearly mapped via a feedforward neural network activated by SwiGLU, outputting an enhanced feature sequence. The core scaling dot product attention formula of the multi-head self-attention mechanism is defined as:

[0131] In the formula, These are the query matrix, key matrix, and value matrix obtained by linearly mapping the input feature sequence, respectively. This represents the transpose of the key matrix. For the feature dimension of the attention head, This is a scaling factor to prevent the dot product result from becoming too large and causing the gradient to vanish.

[0132] Step S223: Reverse reconstruct the enhanced feature sequence through the decoder path to obtain the fused features.

[0133] In this embodiment, the decoder path is a network structure in the deep reconstruction network used to progressively restore the enhanced feature sequence to the original input resolution. Specifically, the decoder path inversely reshapes the enhanced feature sequence into a three-dimensional tensor dimension and uses three-dimensional transposed convolution or three-dimensional trilinear interpolation for progressive upsampling, gradually restoring the resolution of the spatial and temporal depth dimensions. During the decoding stage, a dense skip connection mechanism is employed to concatenate and fuse the upsampled features of the current layer with the historical node features of the same resolution layer and all shallower layers in the encoder path along the channel dimension. This efficiently coordinates the global features extracted by the deep three-dimensional global attention module with the shallow, fine local features, resulting in fused features after restoring the original input resolution.

[0134] The fusion feature is a comprehensive feature tensor output by the decoder path, which integrates deep global semantic information and shallow local detail information. Specifically, the fusion feature simultaneously includes multi-scale local features from the encoder path and global correlation features from the 3D global attention module in the channel dimension, providing rich feature representations for subsequent depth map reconstruction and confidence evaluation.

[0135] Step S224: Input the fused features into the depth map reconstruction branch to obtain a three-dimensional depth distance map of the detected target; wherein, the depth map reconstruction branch is used to independently perform probability normalization along the time depth axis, extract the temporal position of the Gaussian distribution target peak from the obtained single photon time distribution probability, and output a three-dimensional depth distance map.

[0136] In this embodiment, the depth map reconstruction branch is the output network head used to generate a 3D depth range map of the detected target from the fused features. Please continue reading. Figure 4 The depth map reconstruction branch compresses and fuses the channel dimension of features using 3D convolution, and then independently performs probability normalization along the time-depth axis. The probability normalization uses the Softmax activation function for spatial coordinates... The time axis index is The characteristic point, its single-photon echo reflection probability The mathematical formula for calculation is:

[0137] In the formula, For the unnormalized tensor output by the depth branch, The total time step is defined. By probability normalization, the time-domain network response is transformed into a single-photon time distribution probability, which can simultaneously characterize the Gaussian peak of the target reflected signal and the gamma distribution component of the scattering noise.

[0138] The temporal position of the Gaussian-distributed target peak is the time axis index position corresponding to the peak value of the actual reflected echo from the target surface in the single-photon time distribution probability. Specifically, in the single-photon time distribution probability, the echo signal formed by reflection from the target surface exhibits a Gaussian distribution. The peak position of this Gaussian distribution peak corresponds to the actual flight time of the laser pulse from emission to its return to the detector after being reflected by the target surface. The depth map reconstruction branch extracts the temporal position of the Gaussian-distributed target peak, multiplies the time axis index by the time sliding window width, multiplies it by the speed of light, and divides by two to calculate the physical distance corresponding to each spatial location point, thereby outputting a three-dimensional depth-distance map.

[0139] Through the complete processing chain of the encoder path extracting multi-scale local prior features, the three-dimensional global attention module forcibly suppressing gamma-distributed backscattering noise under the global view, the decoder path fusing global and local features, and the depth map reconstruction branch performing probability normalization along the time-depth axis and extracting Gaussian target peaks, the depth reconstruction network can robustly recover high-precision three-dimensional depth and distance information of the detected target from three-dimensional photon count data that is severely contaminated by scattering noise in the harsh environment of haze with extremely low signal-to-noise ratio.

[0140] Please see Figure 5 , Figure 5 The end-to-end processing performance of the system is displayed from top to bottom in environments ranging from light fog to dense fog with extremely low signal-to-noise ratio. For example... Figure 5 As shown in the original depth map extracted using traditional peak detection, in complex hazy environments, the data is filled with discrete and strong backscattering noise caused by aerosol particles, resulting in a large number of noise points or even complete failure in the depth map extracted by traditional methods. After processing by the feature reconstruction model combining multi-scale feature extraction and global attention as described in Embodiment 1 of this invention, the algorithm successfully suppressed the interference noise and output a stable and smooth high-quality target depth-distance map, the final effect of which is shown in the figure. Figure 6 The fog-penetrating reconstruction diagram is shown in the figure. Experimental results verify that the method and system proposed in this invention can effectively overcome the jitter bottleneck of high-capacity transmission and achieve accurate extraction and high-precision 3D reconstruction of the target contour.

[0141] In some embodiments, please refer to Figure 4 In part (c) of the above, in step S222, an attention mechanism is added to the high-order spatiotemporal feature tensor through the three-dimensional global attention module to obtain an enhanced feature sequence, which may include the following steps.

[0142] Step S2221: Perform three-dimensional block labeling on the input deepest high-order spatiotemporal feature tensor and convert it into a one-dimensional feature vector sequence.

[0143] In this embodiment, 3D block tokenization is a preprocessing operation used to transform a high-dimensional 3D feature tensor into a sequence of one-dimensional feature vectors suitable for attention mechanisms. Specifically, to avoid overloading the GPU memory by directly calculating global attention on the entire high-order spatiotemporal feature tensor, 3D block tokenization is performed only on the deepest high-order spatiotemporal feature tensor. According to a preset 3D block size, the high-order spatiotemporal feature tensor is non-overlappingly partitioned in both spatial (height and width) and temporal depth dimensions. All elements within each 3D block are then subjected to mean pooling to reduce dimensionality, resulting in a feature vector. Subsequently, the feature vectors corresponding to all 3D blocks are flattened and expanded in both spatial and temporal dimensions, transforming them into a one-dimensional spatial-temporal feature vector sequence. The 3D block size can be set according to the spatial and temporal depth resolution of the high-order spatiotemporal feature tensor and the GPU memory capacity to achieve a balance between feature granularity and computational efficiency.

[0144] A feature vector sequence is a set of one-dimensional vectors arranged in spatial and temporal order, obtained after three-dimensional block tokenization. Specifically, each element in the feature vector sequence corresponds to a three-dimensional block region in the original high-order spatiotemporal feature tensor, carrying the aggregated spatiotemporal feature information within that local region. The length of the feature vector sequence is equal to the total number of three-dimensional blocks, and the dimension of each feature vector is determined by the size of the three-dimensional block and the number of original feature channels.

[0145] Step S2222: Add three-dimensional positional encoding to the feature vector sequence to obtain the target feature vector sequence.

[0146] In this embodiment, 3D position encoding is an encoded vector used to inject the corresponding 3D spatial and temporal position information into each feature vector in the feature vector sequence. Specifically, since the attention mechanism itself does not have the ability to perceive the positional relationship between elements in the input sequence, the spatial adjacency and temporal order relationships between elements in the feature vector sequence will be lost after the 3D block labeling and flattening operation.

[0147] Therefore, a high-dimensional encoding vector corresponding to its original 3D position is added to each feature vector in the feature vector sequence to obtain the target feature vector sequence. The 3D position encoding can be generated using a sine-cosine position encoding function, whose encoding dimension is consistent with the dimension of the feature vector. It is fused with the feature vector by element-wise addition, so that each feature vector can carry its unique position identifier in 3D spacetime when participating in subsequent attention calculations.

[0148] The target feature vector sequence is a sequence of feature vectors after being injected with 3D positional encoding. Specifically, each feature vector in the target feature vector sequence contains two parts of information: first, local spatiotemporal semantic features from a higher-order spatiotemporal feature tensor; and second, a 3D positional encoding representing its position in the original 3D space. Therefore, the multi-head self-attention mechanism can comprehensively consider both the similarity of feature semantics and the proximity of spatial location when calculating association weights.

[0149] Step S2223: Calculate the correlation weight matrix of the target feature vector sequence using a multi-head self-attention mechanism to suppress gamma-distributed backscattering noise caused by haze aerosols in a global view.

[0150] In this embodiment, the association weight matrix is ​​a weight matrix calculated by a multi-head self-attention mechanism, used to characterize the association strength between any two feature vectors in the target feature vector sequence. Specifically, the target feature vector sequence is linearly mapped to a query matrix Q, a key matrix K, and a value matrix V, respectively, and the attention weight matrix is ​​calculated using the scaled dot product attention formula. Under root mean square normalization stability, each element in the association weight matrix represents the degree of attention one feature vector gives to another feature vector.

[0151] Backscattering noise caused by haze aerosols exhibits specific gamma distribution statistical characteristics in the three-dimensional photon counting matrix, with a wide tail shape in the time depth dimension, while the target reflection signal exhibits Gaussian distribution statistical characteristics, with a relatively concentrated distribution in the time depth dimension.

[0152] The multi-head self-attention mechanism learns and identifies the differences between two statistical distribution patterns by calculating the global correlation weights of nodes across the entire spatiotemporal map. For target signal feature vectors exhibiting a Gaussian distribution, the self-attention mechanism assigns higher correlation weights; for scattering noise feature vectors exhibiting a gamma distribution, it assigns lower correlation weights. Thus, it forcibly suppresses gamma-distributed backscattering noise caused by haze aerosols within a global perspective.

[0153] Step S2224: Multiply the correlation weight matrix with the feature vector sequence to obtain the enhanced feature sequence.

[0154] In this embodiment, the enhanced feature sequence is a new feature vector sequence obtained by multiplying the correlation weight matrix with the corresponding value vectors of the target feature vector sequence. Specifically, the value vectors are weighted and aggregated through the correlation weight matrix, and each output feature vector integrates the information of all input feature vectors in the global scope. Among them, feature vectors with high correlation weights to the current feature vector contribute more information, while feature vectors with low correlation weights contribute less information.

[0155] Thus, the features corresponding to the target Gaussian-distributed echo signal in the enhanced feature sequence are enhanced, while the features corresponding to the gamma-distributed scattering noise are effectively suppressed. Subsequently, the enhanced feature sequence is nonlinearly mapped via a feedforward neural network based on SwiGLU activation to further enrich the expressive power of the features, and finally outputs the enhanced feature sequence for subsequent processing in the decoder path.

[0156] Through the complete processing chain of 3D block marking, 3D position encoding injection, multi-head self-attention global association calculation, and feature weighted aggregation, the 3D global attention module can realize global information interaction across spatiotemporal nodes in the whole map under the constraint of limited video memory. This allows it to effectively distinguish and suppress gamma-distributed backscattering noise caused by haze aerosols in the global view, enhance the feature expression of Gaussian-distributed echo signals reflected from the target surface, and provide high-quality enhanced feature sequences for the decoder path.

[0157] In some embodiments, step S223, in which the enhanced feature sequence is reverse reconstructed through the decoder path to obtain fused features, may include the following steps.

[0158] Step S2231: Following the reverse order of the three-dimensional block tokenization, the enhanced feature sequence is reverse-reshaped into a three-dimensional feature tensor with the same shape as the higher-order spatiotemporal feature tensor.

[0159] In this embodiment, the decoder path receives the enhanced feature sequence output by the 3D global attention module and performs the following reverse reconstruction process to obtain the fused features.

[0160] The reverse process of 3D block tokenization is the inverse operation relative to the forward process of transforming a high-order spatiotemporal feature tensor into a sequence of feature vectors. Specifically, the forward process of 3D block tokenization involves non-overlapping segmentation of the high-order spatiotemporal feature tensor according to a preset 3D block size, dimensionality reduction via mean pooling, and flattening of the spatial and temporal dimensions to obtain a one-dimensional feature vector sequence. Correspondingly, the reverse process of 3D block tokenization involves filling each feature vector in the enhanced feature sequence back into the corresponding 3D block region of the 3D tensor according to its corresponding position in the original 3D space, thereby reshaping it into a 3D feature tensor with a shape consistent with the high-order spatiotemporal feature tensor.

[0161] Since the length of the enhanced feature sequence, the dimension of each feature vector, and the size of the 3D block are all determined during the forward block tokenization process, the reverse reshaping process can completely recover the spatial dimension, temporal depth dimension, and feature channel dimension of the 3D feature tensor.

[0162] The three-dimensional feature tensor is a multidimensional feature representation with a three-dimensional spatial structure obtained by inversely reshaping the enhanced feature sequence. Specifically, the resolution of the three-dimensional feature tensor in the spatial and temporal depth dimensions is consistent with that of the higher-order spatiotemporal feature tensor, that is, it is at the lowest resolution level after multiple downsampling at the deepest layer of the encoder path. In the feature channel dimension, the three-dimensional feature tensor carries feature information enhanced by the global attention mechanism, in which the features of the target Gaussian distributed echo signal are enhanced, and the features of gamma distributed scattering noise are effectively suppressed.

[0163] Step S2232: Using 3D transposed convolution or 3D trilinear interpolation, the 3D feature tensor is upsampled step by step to gradually restore the resolution of the spatial dimension and the temporal depth dimension.

[0164] In this embodiment, 3D transposed convolution is a convolutional operation used to perform learnable upsampling on a 3D feature tensor. Specifically, 3D transposed convolution learns a set of trainable kernel weights to expand each element of the input feature tensor into the local 3D neighborhood of the output feature tensor according to the kernel weights, thereby simultaneously improving resolution in both spatial and temporal depth dimensions. The upsampling ratio of 3D transposed convolution is determined by the kernel size, stride, and padding parameters, enabling end-to-end trainable mapping from low-resolution features to high-resolution features.

[0165] 3D trilinear interpolation is an interpolation algorithm used for non-parametric upsampling of 3D feature tensors. Specifically, for each target location in the high-resolution feature tensor to be generated, 3D trilinear interpolation performs linear interpolation along the spatial height, spatial width, and temporal depth directions in the 3D grid of the low-resolution feature tensor. The weighted average of the distances between the target location and its eight adjacent grid points is calculated as the interpolation result for that location. Compared with 3D transposed convolution, 3D trilinear interpolation does not require learning additional convolution kernel parameters, resulting in higher computational efficiency. It is suitable for applications with moderate resolution restoration accuracy requirements but high real-time requirements.

[0166] Stepwise upsampling is the process of gradually restoring the resolution of a 3D feature tensor from its deepest low-resolution state to the original input resolution. Specifically, the decoder path has the same number of upsampling levels as the encoder path's spatial downsampling levels. Each upsampling level increases the spatial and temporal depth resolution of the current feature tensor to the resolution level of the corresponding downsampling level in the encoder path. By upsampling stepwise rather than upsampling to the original resolution all at once, the feature recovery process after each upsampling level is smoother and more controllable, which is beneficial for preserving and fusing detailed information at different scales.

[0167] Step S2233: After upsampling at each level, the upsampled features of the current level are spliced ​​and fused with the historical node features of the same resolution level and all shallower levels in the encoder path in the channel dimension through a dense skip connection mechanism until the original input resolution is restored, and the fused features are obtained.

[0168] In this embodiment, the dense skip connection mechanism is a connection method used to fully fuse the upsampled features of the decoder path with the historical node features of multiple levels in the encoder path. Specifically, after each level of upsampling, the decoder path not only skips the upsampled features of the current level with the historical node features of the same resolution level in the encoder path, but also splices and fuses the upsampled features of the current level with all historical node features of the encoder path that are shallower and have higher resolution than the current level in the channel dimension.

[0169] Historical node features are the intermediate feature tensors output by each spatial downsampling level during the forward propagation of the encoder path. Specifically, due to the densely nested architecture of the encoder path, each spatial downsampling level not only passes features to the next level but also saves its output intermediate feature tensors as historical node features. These historical node features contain spatial prior information with different levels of semantic abstraction from shallow to deep layers: shallower-level historical node features retain more spatial details and edge information, while deeper-level historical node features contain more abstract target semantic features. Through a dense skip connection mechanism, the current upsampling features of the decoder path can simultaneously acquire directly corresponding features from the same resolution level and progressively supplementary features from all shallower levels, thus forming a dense stitching of multi-scale features in the channel dimension.

[0170] Channel-dimensional concatenation and fusion is an operation that concatenates multiple feature tensors along the channel dimension. Specifically, for the upsampled features of the current layer and the features of multiple historical nodes participating in skip connections, the decoder path concatenates them sequentially along the channel dimension while maintaining consistency in the spatial and temporal depth dimensions of each feature tensor, forming a fused feature tensor with the number of channels equal to the sum of the number of channels of each input feature tensor. Subsequently, convolutional operations are used to perform channel compression and information integration on the concatenated fused feature tensor to achieve efficient collaboration between deep global features and shallow local features. This continues until the last upsampled layer is restored to the same spatial and temporal depth resolution as the original input 3D photon counting matrix, outputting the fused feature.

[0171] Through the complete decoding process of reverse reshaping, step-by-step upsampling, and dense skip connections described above, the decoder path can gradually restore the enhanced feature sequence output by the 3D global attention module from the deepest low-resolution abstract representation to the original input resolution. In the process of restoration, it fully integrates the multi-scale local prior features of each level of the encoder path, so that the final output fused features have both global scattering noise suppression capability and local spatial detail preservation capability, providing high-quality comprehensive feature representation for the subsequent depth map reconstruction branch and confidence evaluation branch.

[0172] In some embodiments, after step S223, the data processing method for single-photon fog-penetrating imaging may further include the following steps.

[0173] Step S225: Input the fused features into the confidence evaluation branch to obtain the confidence evaluation matrix of the detected target; wherein, the confidence evaluation branch is used to map the output of the confidence matrix consistent with the spatial resolution using the Sigmoid activation function, so as to quantify the physical reliability of the depth value of each pixel.

[0174] In this embodiment, please continue to refer to Figure 4 In this embodiment, after the decoder outputs the fused features, the depth reconstruction network adopts a dual-branch output structure, setting up a depth map reconstruction branch and a confidence evaluation branch in parallel. The fused features are simultaneously input to these two independent output network heads, which are used to generate a 3D depth distance map and a confidence evaluation matrix, respectively.

[0175] The confidence evaluation branch is the output network head in the deep reconstruction network used to evaluate the physical reliability of the depth values ​​of each pixel. Specifically, the confidence evaluation branch shares the same fused features as the depth map reconstruction branch as input, but uses independent network parameters and activation functions to extract feature information related to the confidence of the depth estimation from the fused features. The parallel output design of the two branches enables the deep reconstruction network to simultaneously provide the depth estimation result and the reliability evaluation of the estimation result, providing richer decision-making basis for downstream applications (such as object recognition, path planning, etc.).

[0176] The Sigmoid activation function is a sigmoid nonlinear activation function used to map input values ​​to the interval (0,1). Specifically, the mathematical expression for the Sigmoid activation function is: ,in, The input value is denoted as . When the input value approaches positive infinity, the output value approaches 1; when the input value approaches negative infinity, the output value approaches 0; when the input value is 0, the output value is 0.5. Using the Sigmoid activation function, the confidence evaluation branch maps the confidence features extracted from the fused features to normalized confidence scalar values, giving the output results an intuitive probabilistic interpretation.

[0177] The confidence matrix is ​​a two-dimensional matrix with the same spatial dimensional resolution as the input 3D photon counting matrix. Specifically, the confidence evaluation branch compresses the channel dimension of the fused features using 3D convolution, then independently applies the Sigmoid activation function to the feature vector at each spatial location for mapping, outputting a two-dimensional confidence matrix with the same resolution as the spatial height and width dimensions. For spatial coordinates... The confidence scalar of the pixels. The mathematical mapping formula is:

[0178] In the formula, This is the raw response value output by the confidence assessment branch. Each element in the confidence matrix ranges from [0,1], with larger values ​​indicating higher physical reliability of the depth estimation result for the corresponding pixel.

[0179] Physical reliability is an evaluation metric used to quantify the consistency between the depth distance values ​​of each pixel output by the deep reconstruction network and the actual physical distance. Specifically, in complex meteorological environments such as fog and haze, the signal-to-noise ratio of the reflected signal from the target surface varies at different spatial locations: for pixels with strong reflected signals (such as flat areas on the target surface directly facing the lidar), the network can accurately distinguish the Gaussian target peak from the scattering noise, and the depth value of this pixel has high physical reliability; for pixels with weak reflected signals (such as target edge areas or areas severely attenuated by dense fog), the Gaussian target peak may be completely submerged by gamma-distributed scattering noise, and the depth value output by the network may have greater uncertainty, resulting in correspondingly lower physical reliability for this pixel.

[0180] In some embodiments, the confidence assessment matrix can be used for depth value screening and filtering in subsequent processing. Specifically, a confidence threshold can be set, and pixels with a confidence level below the threshold are marked as unreliable and filtered out, while only the depth values ​​corresponding to pixels with a confidence level above the threshold are retained for constructing the final 3D point cloud image. This effectively suppresses uncertain depth points in the depth map caused by scattering noise, improving the reliability and usability of the 3D imaging results.

[0181] Through the aforementioned confidence evaluation branch, the depth reconstruction network can not only output a 3D depth-distance map of the detected target, but also simultaneously output a confidence matrix consistent with the spatial resolution to quantify the physical reliability of the depth values ​​of each pixel. This dual-branch output structure enables the single-photon fog-penetrating imaging system to self-evaluate the quality of its depth estimation results in practical engineering applications, providing more comprehensive information support for subsequent target detection, 3D reconstruction, and other tasks.

[0182] Please see Figure 6One embodiment of the present invention can provide a data processing system for single-photon fog-penetrating imaging, the system comprising an airborne terminal system and a ground terminal system.

[0183] The airborne terminal system includes a data acquisition module, a first data processing module, and an airborne wireless communication module.

[0184] The data acquisition module is used to acquire the raw echo signal of the single-photon detection system to the detection target, and to converge the raw echo signal to obtain multiple high-frequency pulse data packets.

[0185] The first data processing module, connected to the data acquisition module, is used to perform anti-jitter buffering and dynamic load reassembly on the high-frequency pulse data packets to obtain the target large packet data, and enqueue it into a pre-set asynchronous circular queue.

[0186] An airborne wireless communication module, connected to the first data processing module, is used to transmit the target large packet data in the asynchronous circular queue to the ground terminal according to a preset sending thread.

[0187] The ground terminal system includes a host computer, which includes a second data processing module and a deep reconstruction module.

[0188] The second data processing module is used to construct a three-dimensional photon counting matrix, including temporal depth and spatial dimensions, for one or more received target large data packets.

[0189] The deep reconstruction module, connected to the second data processing module, is used to input the three-dimensional photon counting matrix into a pre-trained deep reconstruction network to obtain a three-dimensional depth distance map of the detected target; wherein, the deep reconstruction network uses multi-scale feature extraction and global attention mechanism to suppress aerosol scattering noise.

[0190] In this embodiment, the data processing system adopts a distributed hardware architecture of airborne and ground-based terminals, including an airborne terminal system and a ground-based terminal system. Specifically, the airborne terminal system undertakes the hardware execution tasks of underlying signal acquisition, buffering and shaping, and flow shaping. The ground-based terminal system, relying on its powerful computing capabilities, is responsible for performing the construction of a three-dimensional photon counting matrix and deep network reconstruction operations, including global attention scattering suppression, thereby achieving end-to-end hardware and software collaboration.

[0191] The data acquisition module is used to acquire the raw echo signal from the single-photon detection system targeting the target under complex weather conditions, and to converge the raw echo signal into multiple high-frequency pulse data packets. Specifically, the data acquisition module may include a lidar data acquisition module. This lidar data acquisition module includes conventional optical and sensing components such as a single-photon detector, and is used to emit laser pulses towards the target and receive photon echo signals reflected from the target surface and scattered by the environment.

[0192] Furthermore, the lidar data acquisition module may also include a high-speed time measurement module and a bus expansion module. The high-speed time measurement module performs high-precision time-to-digital conversion on multi-channel echo signals to obtain photon arrival timestamp information; the bus expansion module aggregates multiple pulse signals or low-level control signals for unified reading and processing by subsequent modules. In a specific example, the high-speed time measurement module can use an HST-4A multi-channel high-speed time-to-digital converter, and the bus expansion module can use a Yabo intelligent USB expansion board.

[0193] The first data processing module is connected to the data acquisition module and is used to perform anti-jitter buffering and dynamic load reassembly on the high-frequency pulse data packets to obtain the target large packet data and enqueue it into a pre-set asynchronous circular queue.

[0194] The airborne wireless communication module is connected to the first data processing module and is used to transmit the target large packet data in the asynchronous circular queue to the ground terminal according to a preset transmission thread. Specifically, the airborne wireless communication module transmits the shaped constant data stream to the ground through a multi-level heterogeneous physical link. The multi-level heterogeneous physical link includes a multi-level physical medium bridge from the underlying serial bus and local area network Ethernet to space radio frequency microwave. In a specific example, the airborne wireless communication module can use an LQ3 image transmission module or other equipment set to AP (mobile terminal) mode as the airborne transmitting terminal.

[0195] The ground terminal system is deployed in the ground data processing center. The ground terminal system includes a host computer. The host computer includes a second data processing module and a deep reconstruction module.

[0196] The second data processing module is used to construct a three-dimensional photon counting matrix, including temporal depth and spatial dimensions, for one or more received target large packet data. Specifically, the second data processing module extracts the photon counting features from the transmitted target large packet data, performs matrix mapping according to the physical correspondence between temporal sliding windows, spatial height, and spatial width, and constructs a three-dimensional photon counting matrix with dimensions B×C×D×H×W, providing a standardized input format for the subsequent deep reconstruction network.

[0197] The deep reconstruction module is connected to the second data processing module and is used to input the three-dimensional photon counting matrix into a pre-trained deep reconstruction network to obtain a three-dimensional depth distance map of the detected target. The deep reconstruction network utilizes multi-scale feature extraction and a global attention mechanism to suppress aerosol scattering noise. Specifically, the deep reconstruction module internally runs the aforementioned deep reconstruction network, which may include an encoder path, a three-dimensional global attention module, and a decoder path. It employs a densely nested architecture and a multi-head self-attention mechanism to forcibly suppress gamma-distributed backscattering noise caused by haze aerosols within a global field of view. Finally, it outputs the three-dimensional depth distance map and a confidence evaluation matrix through dual-branch output heads. In a specific example, the host computer can be a personal computer, workstation, or dedicated industrial control computer equipped with a high-performance graphics processing unit (GPU).

[0198] Through the distributed hardware architecture of the airborne terminal system and the ground terminal system and the collaborative work between their modules, the data processing system provided in this application embodiment can solve the two major technical problems of transmission link congestion and haze scattering noise without increasing the overhead of airborne hardware resources, and achieve highly reliable single-photon fog-penetrating imaging in harsh environments with extremely low signal-to-noise ratio.

[0199] In some embodiments, the first data processing module includes a network peripheral server and an airborne computer. The network peripheral server is connected to the data acquisition module and is used to achieve transparent pass-through and virtualization bridging of the underlying physical interface based on bus virtualization technology, so as to access the high-frequency pulse data packets into the local area network. The input and output terminals of the airborne computer are respectively connected to the network peripheral server and the airborne wireless communication module, and are used to receive the high-frequency pulse data packets passing through the local area network, and to perform anti-jitter buffering and dynamic load reassembly on the high-frequency pulse data at the airborne gateway layer to generate target large packet data, and send it into a pre-set asynchronous circular queue.

[0200] Please continue reading. Figure 6 In this embodiment, the first data processing module consists of a two-level hardware structure of a network peripheral server and an airborne computer. The two work together to complete the complete data processing link from the underlying physical interface data acquisition to the airborne gateway layer traffic shaping.

[0201] A network peripheral server is a hardware device used to implement protocol conversion and virtualization bridging between the underlying physical interface and the local area network (LAN). Specifically, the input end of the network peripheral server is physically connected to the lidar data acquisition module in the data acquisition module, and its output end is connected to the onboard computer through the LAN interface.

[0202] Bus virtualization technology is used to map physically local high-speed serial bus devices into network-reachable virtual devices at the logical level. Specifically, a network peripheral server, based on bus virtualization technology, encapsulates and converts the data communication on the high-speed serial bus (such as a USB bus) to which the LiDAR data acquisition module is connected, enabling the underlying physical interfaces that were originally only accessible on the local host to be accessed by a remote host over the network. In a specific example, the network peripheral server can use USB virtualization technology to achieve the above functionality.

[0203] Transparent pass-through is a real-time forwarding method where network peripheral servers forward data from the underlying physical interface without modification or caching. Specifically, transparent pass-through means that when a network peripheral server connects high-speed serial bus data to a local area network, it does not parse, modify, or filter the data content; it only performs necessary protocol encapsulation and forwarding, thereby ensuring bit-level consistency of data during transmission from the LiDAR data acquisition module to the onboard computer.

[0204] Virtualization bridging is a logical connection channel established between the underlying physical interface of a network peripheral server and the local area network (LAN). Specifically, virtualization bridging enables the host computer in the ground terminal system to directly access the optical sensing or low-level control components within the LiDAR data acquisition module at the logical level, overcoming the physical isolation between the airborne and ground terminals via a virtualized link. This real-time direct access capability across physical isolation is crucial for applications such as system debugging, parameter configuration, and status monitoring. In a specific example, the network peripheral server can utilize the Silex-510 device, which supports synchronous transmission with extremely high real-time requirements, ensuring continuous and uninterrupted data transmission.

[0205] A local area network (LAN) is a local data communication network between a network peripheral server and an airborne computer. Specifically, by connecting high-frequency pulse data packets to the LAN, the high-speed serial bus data generated by the data acquisition module is converted into a standard network data packet format, enabling the airborne computer to receive and process data through a standard network protocol stack interface. This reduces the coupling between data acquisition and data processing, and improves the modularity and scalability of the system.

[0206] The airborne computer is a hardware computing platform used to execute core data processing logic at the airborne gateway layer. Specifically, the input end of the airborne computer is connected to the network peripheral server through a network interface to receive high-frequency pulse data packets transmitted through the local area network; the output end of the airborne computer is connected to the airborne wireless communication module through a network interface to transmit the target large packet data, which has been processed by anti-jitter buffering and dynamic load reassembly, to the ground end according to a preset transmission thread.

[0207] The airborne gateway layer is a data processing software layer running inside the airborne computer, used to execute core processing logic such as anti-jitter buffering, dynamic load reassembly, and proactive traffic shaping. Specifically, during the initialization phase, the airborne gateway layer allocates and locks contiguous physical memory and establishes an asynchronous circular queue protected by a mutex lock. During the operation phase, it captures high-frequency pulse data packets through independent parallel receiving threads, extracts the effective payload, and performs fixed-length splicing. When the length of the spliced ​​accumulated payload reaches a preset output load threshold, it encapsulates and generates the target large packet data and writes it to the asynchronous circular queue. At the same time, it cyclically extracts the target large packet data from the asynchronous circular queue through independent parallel sending threads, dynamically calculates the microsecond-level safe sending waiting time interval based on the preset target throughput bit rate and the target number of packets per second, and calls a blocking sleep function to suspend the sending thread within this time interval, thereby realizing the transformation from sudden high-frequency pulse data surges to a smooth and constant output data stream.

[0208] Through the two-tier hardware architecture and functional division of the network peripheral server and airborne computer described above, the first data processing module separates the data acquisition of the underlying physical interface from the data shaping processing of the airborne gateway layer at the hardware level. This allows the data acquisition stage to focus on ensuring the real-time performance and integrity of the data, while the data processing stage focuses on executing computationally intensive buffer shaping and traffic scheduling algorithms. In this way, reliable end-to-end data preprocessing and shaping transmission can be achieved without increasing the processing burden of a single hardware node.

[0209] In some embodiments, the ground terminal system further includes a ground wireless communication module. The ground wireless communication module is used to establish a cross-terminal wireless data link with the airborne wireless communication module and to receive target large packet data sent by the airborne terminal system. The ground wireless communication module is connected to the host computer and is used to transmit the target large packet data to the host computer.

[0210] Please see Figure 6 In this embodiment, the ground terminal system consists of a ground wireless communication module and a host computer, which are interconnected to form a complete ground terminal data reception and processing link.

[0211] A terrestrial wireless communication module is a communication device used to establish a cross-end wireless data link with an airborne wireless communication module in an airborne terminal system. Specifically, the terrestrial wireless communication module and the airborne wireless communication module communicate wirelessly via air interface electromagnetic waves, forming a data transmission bridge between the airborne terminal system and the terrestrial terminal system.

[0212] A cross-end wireless data link is an end-to-end wireless communication connection established between an airborne wireless communication module and a ground-based wireless communication module. Specifically, the cross-end wireless data link can transmit signals based on the radio frequency microwave band, and its transmission distance, bandwidth, and anti-interference capability depend on the performance parameters of the wireless communication module used. Relying on microsecond-level active traffic shaping performed by the airborne terminal system, the target large packet data transmitted via the cross-end wireless data link arrives at the ground end at a smooth and constant rate. This ensures that the transmission process of the cross-end wireless data link strictly matches the throughput capacity of the wireless communication module, effectively avoiding link congestion and packet loss caused by data bursts.

[0213] In some embodiments, the operating mode of the ground wireless communication module is configured to correspond to the operating mode of the airborne wireless communication module. Specifically, when the airborne wireless communication module is set to AP (Mobile Terminal) mode as an airborne transmitting terminal, the ground wireless communication module is correspondingly set to STA (Station Terminal) mode as a ground receiving terminal. The ground wireless communication module in STA mode can actively search for and connect to the wireless network established by the airborne wireless communication module in AP mode, thereby maintaining a stable connection of the cross-end wireless data link during UAV flight operations. In a specific example, the ground wireless communication module can employ devices such as the LQ3 image transmission module, which provides stable air interface transmission performance and is suitable for long-distance data downlink applications of UAVs.

[0214] The reception and transmission of target large packet data is the core function of the ground wireless communication module. Specifically, the ground wireless communication module receives target large packet data sent by the airborne wireless communication module through a cross-end wireless data link. Because the target large packet data has already undergone anti-jitter buffering, dynamic load reassembly, and microsecond-level active flow shaping at the airborne end, it exhibits a smooth and constant rate characteristic when it arrives at the ground wireless communication module. This avoids overflow of the ground-end receiving buffer due to data bursts, ensuring the integrity and timing accuracy of the target large packet data.

[0215] The terrestrial wireless communication module connects to the host computer via wired or wireless means. Specifically, after receiving the target large packet data, the terrestrial wireless communication module transmits it to the host computer through the connected communication interface. The host computer's internal second data processing module and deep reconstruction module then perform subsequent processing on the target large packet data. The communication interface can be selected according to the host computer's hardware configuration and interface type, such as a USB interface, Ethernet interface, or serial communication interface, etc. This embodiment does not impose specific limitations here.

[0216] Through the hardware architecture and functional division of the above-mentioned ground wireless communication module and host computer, the ground terminal system separates the wireless data reception function and the deep reconstruction calculation function at the hardware level. This allows the wireless communication module to focus on ensuring the integrity and real-time performance of data reception, while the host computer focuses on performing computationally intensive matrix reconstruction and deep learning inference tasks. In this way, end-to-end high-reliability data reception and high-precision 3D imaging can be achieved without increasing the processing burden of a single hardware node.

[0217] It is understood that the specific examples in this document are only intended to help those skilled in the art better understand the embodiments described herein, and are not intended to limit the scope of the invention.

[0218] It is understood that in the various embodiments described in this specification, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments described in this specification.

[0219] It is understood that the various implementation methods described in this specification can be implemented individually or in combination, and the implementation methods in this specification are not limited in this respect.

[0220] Unless otherwise stated, all technical and scientific terms used in the embodiments of this specification have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this specification. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items. The singular forms "a," "the," and "the" as used in the embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0221] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this specification.

[0222] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the aforementioned method implementations, and will not be repeated here.

[0223] In the several embodiments provided in this specification, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0224] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0225] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0226] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of this specification, in essence, or the parts that contribute to the prior art, or parts of the technical solutions, can be embodied in the form of software products. These computer software products are stored in a storage medium and include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this specification. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0227] The above description is merely a specific embodiment of this specification, but the scope of protection of this invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this specification should be included within the scope of protection of this specification. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A data processing method for single-photon fog-penetrating imaging, characterized in that, The method is applied to a data processing system including airborne and ground-based terminals, and the method includes: The onboard terminal acquires the raw echo signal of the single-photon detection system to the target and aggregates the raw echo signal to obtain multiple high-frequency pulse data packets. The high-frequency pulse data packets are subjected to anti-jitter buffering and dynamic load reassembly to obtain the target large packet data, which is then enqueued into a pre-set asynchronous circular queue. The target large packet data in the asynchronous circular queue is transmitted to the ground terminal according to the preset sending thread; Using the ground terminal, a three-dimensional photon counting matrix, including temporal depth and spatial dimensions, is constructed for one or more target large data packets received. The three-dimensional photon counting matrix is ​​input into a pre-trained deep reconstruction network to obtain a three-dimensional depth distance map of the detected target; wherein, the deep reconstruction network uses multi-scale feature extraction and global attention mechanism to suppress aerosol scattering noise; Specifically, transmitting the target large packet data in the asynchronous circular queue to the ground terminal according to a preset sending thread includes: The target large packet data is extracted from the asynchronous circular queue in a loop using independent and parallel sending threads; Based on a pre-set target throughput bit rate and target packets per second, combined with the actual byte length of the currently extracted target large packet data, the microsecond-level secure transmission waiting time interval required for the current transmission is dynamically calculated. This includes: calculating a first theoretical transmission time limited by bandwidth based on the actual byte length of the target large packet data and the target throughput bit rate; calculating a second theoretical transmission time limited by packet rate based on the target packets per second; determining the maximum value between the first theoretical transmission time and the second theoretical transmission time as the expected transmission time interval for the target large packet data; obtaining the system elapsed time from the completion of the last transmission to the current moment; and determining the microsecond-level secure transmission waiting time interval based on the difference between the expected transmission time interval and the system elapsed time. During the microsecond-level safe transmission waiting time interval, a blocking sleep function is invoked to suspend the transmission thread.

2. The method according to claim 1, characterized in that, The high-frequency pulse data packets are subjected to anti-jitter buffering and dynamic payload reassembly to obtain the target large packet data, which is then enqueued into a pre-set asynchronous circular queue, including: Allocate and lock a contiguous block of physical memory, and establish the asynchronous circular queue protected by a mutex lock on the physical memory; The high-frequency pulse data packets are captured by independent and parallel receiving threads, the payload of the data packets is extracted in real time, and the payload is spliced ​​into fixed lengths in the physical memory. When the accumulated load length reaches the preset output load threshold, the spliced ​​data is encapsulated into the target large packet data and written into the asynchronous circular queue.

3. The method according to claim 2, characterized in that, The airborne terminal also includes a runtime state monitoring thread and an out-of-band serial port control thread, independent of the sending thread and the receiving thread; the method further includes: Without system shutdown and without interruption of data flow, the out-of-band serial port control thread receives external configuration adjustment instructions and uses the system mutex lock to safely update the parameter variables in the runtime control structure, so that the sending thread applies the updated parameter variables in the next processing cycle. The parameter variables include the target throughput bit rate and / or the output load threshold.

4. The method according to claim 1, characterized in that, For one or more target large data packets received, a three-dimensional photon counting matrix including temporal and spatial dimensions is constructed, including: Extract photon count features from one or more target large packet data transmitted through the tunnel, and perform matrix mapping according to the physical correspondence between time sliding window, spatial height and spatial width; A three-dimensional photon counting matrix with dimensions B×C×D×H×W is constructed; where B is the batch size, C is the number of feature channels, D is the temporal depth dimension, and H and W are the height and width dimensions of the spatial image, respectively.

5. The method according to claim 4, characterized in that, The deep reconstruction network includes an encoder path, a 3D global attention module, and a decoder path. The 3D photon counting matrix is ​​input into the pre-trained deep reconstruction network to obtain a 3D depth distance map of the detected target, including: Through the encoder path, the three-dimensional photon counting matrix is ​​controlled to flow through multiple spatial downsampling levels to extract local features, and a high-order spatiotemporal feature tensor is output at the deepest layer. The three-dimensional global attention module adds an attention mechanism to the high-order spatiotemporal feature tensor to obtain an enhanced feature sequence; wherein, the three-dimensional global attention module is used to suppress gamma distribution backscattering noise caused by haze aerosols. The enhanced feature sequence is reconstructed in reverse using the decoder path to obtain the fused features; The fused features are input into the depth map reconstruction branch to obtain a three-dimensional depth distance map of the detected target; wherein, the depth map reconstruction branch is used to independently perform probability normalization along the time depth axis, extract the temporal position of the Gaussian distribution target peak from the obtained single photon time distribution probability, and output a three-dimensional depth distance map.

6. The method according to claim 5, characterized in that, The three-dimensional global attention module adds an attention mechanism to the high-order spatiotemporal feature tensor to obtain an enhanced feature sequence, including: The input deepest high-order spatiotemporal feature tensor is divided into three-dimensional blocks and labeled, and then transformed into a one-dimensional feature vector sequence. Add three-dimensional positional encoding to the feature vector sequence to obtain the target feature vector sequence; The correlation weight matrix of the target feature vector sequence is calculated using a multi-head self-attention mechanism to suppress gamma-distributed backscattering noise caused by haze aerosols in a global view. Multiplying the correlation weight matrix by the feature vector sequence yields the enhanced feature sequence.

7. The method according to claim 6, characterized in that, The enhanced feature sequence is reconstructed in reverse through the decoder path to obtain fused features, including: Following the reverse order of the three-dimensional block tokenization, the enhanced feature sequence is reverse-reshaped into a three-dimensional feature tensor with the same shape as the higher-order spatiotemporal feature tensor; By using 3D transposed convolution or 3D trilinear interpolation, the 3D feature tensor is upsampled step by step to gradually restore the resolution of the spatial dimension and the temporal depth dimension; After each upsampling stage, the upsampled features of the current level are spliced ​​and fused with the historical node features of the same resolution level and all shallower levels in the encoder path through a dense skip connection mechanism in the channel dimension until the original input resolution is restored, thus obtaining the fused features.

8. The method according to claim 5, characterized in that, After the step of reconstructing the enhanced feature sequence in reverse through the decoder path to obtain the fused features, the following steps are included: The fused features are input into the confidence evaluation branch to obtain the confidence evaluation matrix of the detected target; wherein, the confidence evaluation branch is used to map the output of the confidence matrix consistent with the spatial resolution using the Sigmoid activation function, so as to quantify the physical reliability of the depth value of each pixel.

9. A data processing system for single-photon fog-penetrating imaging, characterized in that, The system is used to implement the data processing method for single-photon fog-penetrating imaging as described in any one of claims 1 to 8, and the system includes an airborne terminal system and a ground terminal system; The airborne terminal system includes: The data acquisition module is used to acquire the raw echo signal of the single-photon detection system to the detection target, and to converge the raw echo signal to obtain multiple high-frequency pulse data packets; The first data processing module, connected to the data acquisition module, is used to perform anti-jitter buffering and dynamic load reassembly on the high-frequency pulse data packets to obtain the target large packet data and enqueue it into a pre-set asynchronous circular queue. An airborne wireless communication module, connected to the first data processing module, is used to transmit the target large packet data in the asynchronous circular queue to the ground terminal according to a preset sending thread; The ground terminal system includes a host computer, which includes: The second data processing module is used to construct a three-dimensional photon counting matrix, including temporal depth and spatial dimensions, for one or more target large data packets received. The deep reconstruction module, connected to the second data processing module, is used to input the three-dimensional photon counting matrix into a pre-trained deep reconstruction network to obtain a three-dimensional depth distance map of the detected target; wherein, the deep reconstruction network uses multi-scale feature extraction and global attention mechanism to suppress aerosol scattering noise.

10. The system according to claim 9, characterized in that, The first data processing module includes: A network peripheral server, connected to the data acquisition module, is used to achieve transparent transmission and virtualization bridging of the underlying physical interface based on bus virtualization technology, and to access the high-frequency pulse data packets into the local area network. An airborne computer, whose input and output terminals are respectively connected to the network peripheral server and the airborne wireless communication module, is used to receive high-frequency pulse data packets passing through the local area network, and to perform anti-jitter buffering and dynamic load reassembly on the high-frequency pulse data at the airborne gateway layer to generate target large packet data, and send it into a pre-set asynchronous circular queue.

11. The system according to claim 9, characterized in that, The ground terminal system also includes: The ground wireless communication module is used to establish a cross-terminal wireless data link with the airborne wireless communication module and to receive target large packet data sent by the airborne terminal system. The ground wireless communication module is connected to the host computer and is used to transmit the target large packet data to the host computer.

Images