Adaptive Target Detection Method and System for TOF Point Cloud

Abstract

An adaptive target detection method for TOF point clouds, the method comprising: receiving TOF point cloud data of a target to be detected collected by a TOF camera; sequentially using the DB-SCAN algorithm and the binary iterative k-means algorithm on the TOF point cloud data to obtain preliminarily denoised point cloud data; using the k-means algorithm for the depth information of the preliminarily denoised point cloud data to obtain the depth range of the distribution of the target to be detected; using a pre-established and trained Gaussian mixture model to cluster the preliminarily denoised point cloud data within the depth range to determine the point cloud data corresponding to different targets to be detected; using the iterative bilateral filtering algorithm to perform local denoising processing on the point cloud data of each target to be detected respectively; and inputting the denoised point cloud data of each target to be detected into a pre-established and trained target detection model respectively to complete the recognition of each target.

Description

Technical Field

[0001] The present invention relates to the technical fields of denoising processing and target detection based on TOF point cloud data, and particularly relates to an adaptive target detection method and system for TOF point cloud. Background Art

[0002] For the target detection task, denoising is a crucial step, and the distribution of effective data directly affects the final detection effect. Existing point cloud denoising algorithms usually require a large number of parameters and thresholds to be set subjectively by humans, and it is difficult to reuse the same set of parameters. When the data volume is huge, appropriate parameters need to be selected manually for each point cloud map, which obviously consumes a large amount of human resources. For example, bilateral filtering, Gaussian filtering, KD-Tree, random sample consensus filtering, etc.

[0003] TOF, (Time of Flight), the time of flight, its basic principle is to continuously emit light pulses (generally invisible light) onto the object to be observed, and then use a sensor to receive the light returned from the object, and directly obtain the distance of the target object by detecting the flight (round trip) time of the light pulse.

[0004] TOF has the characteristics of high frame rate (the frame rate range includes 10 - 120fps), fast phase demodulation speed (<1ms), and direct and convenient acquisition of three-dimensional information. It has the advantages that the measurement accuracy will not decrease with the increase of the measurement distance, the measurement error is basically fixed within the entire measurement range, and the TOF depth camera also has strong anti-interference ability. Therefore, in occasions where the measurement distance requirement is relatively far (such as unmanned driving, space unmanned robot detection), the TOF depth camera has very obvious advantages.

[0005] However, compared with other ways of obtaining point cloud data (such as binocular stereo vision, structured light, etc.), it is more susceptible to multi-path interference generated by the environment, the reflection characteristics of the target, etc. The number of noise points in the directly collected data is large and the density is high. When directly using the above traditional filtering methods, a large number of parameter thresholds need to be set manually, and the effect is not satisfactory, and the signal-to-noise ratio is still very large, seriously affecting the final recognition result. Summary of the Invention

[0006] The filtering algorithms of the prior art rely too much on manual subjective parameter setting for the point cloud data directly collected by the TOF camera, and the filtering effect is not satisfactory. The purpose of the present invention is to overcome the above defects of the prior art and propose an adaptive target detection system for TOF point cloud.

[0007] To achieve the above purpose, the present invention proposes an adaptive target detection method for TOF point cloud, and the method includes:

[0008] Receive the TOF point cloud data of the target to be detected collected by the TOF camera;

[0009] Apply the DB-SCAN algorithm and the binary iterative k-means algorithm to the TOF point cloud data in sequence to obtain the preliminarily denoised point cloud data;

[0010] Apply the k-means algorithm to the depth information of the preliminarily denoised point cloud data to obtain the depth range of the distribution of the target to be detected;

[0011] Use the pre-established and trained Gaussian mixture model to cluster the preliminarily denoised point cloud data within the depth range to determine the point cloud data corresponding to different targets to be detected;

[0012] Use the iterative bilateral filtering algorithm to perform local denoising processing on the point cloud data of each target to be detected respectively;

[0013] Input the denoised point cloud data of each target to be detected into the pre-established and trained target detection model respectively to complete the recognition of each target.

[0014] As an improvement of the above method, by applying the DB-SCAN algorithm to the TOF point cloud data, find the outlier points generated due to the influence of multipath reflection and eliminate them.

[0015] As an improvement of the above method, apply the binary iterative k-means algorithm to the point cloud data after eliminating the outlier points, specifically including:

[0016] Take the direction perpendicular to the ground as the measurement dimension, apply the k-means algorithm to perform binary clustering on the point cloud data after eliminating the outlier points to obtain two types of point cloud data;

[0017] Select the type of point cloud data with more data points from them, apply the k-means algorithm for binary clustering again, repeat this step until the set number of rounds is reached to obtain the segmentation points that meet the preset requirements, find the ground noise points to be eliminated and eliminate them.

[0018] As an improvement of the above method, when applying the k-means algorithm to the preliminarily denoised point cloud data and combining the depth information in the point cloud data to obtain the depth range of the distribution of the target to be detected; specifically including:

[0019] Apply the PCA algorithm to the point cloud data within each target to be detected area to obtain the optimal mapping hyperplane, so that the variance of the distribution of the point cloud data within the target to be detected area is the largest, perform dimensionality reduction mapping on the point cloud data, and record the corresponding relationship of the data points before and after the dimensionality reduction mapping at the same time;

[0020] According to the preset classification number k, a k-means clustering is performed on the reduced-dimensional point cloud data to determine the data points of the target to be detected, and the corresponding original point cloud data is found according to the data relationship before and after the dimensionality reduction mapping. This part of the data is retained, and other data points are eliminated. The depth range of the target distribution to be detected is obtained by combining the depth information in the point cloud data.

[0021] As an improvement of the above method, the target detection model is a PointNet neural network, the input is the point cloud data of the target to be detected after denoising, and the output is the target recognition result.

[0022] As an improvement of the above method, before receiving the TOF point cloud data of the target to be detected collected by the TOF camera, the method further includes: pre-establishing a blocking queue for each TOF camera for caching the TOF point cloud data; the blocking queue supports first-in-first-out read and write operations, specifically including:

[0023] Through the write operation, the TOF point cloud data of the target to be detected is stored in the blocking queue in time sequence, and through the read operation, the TOF point cloud data is obtained from the blocking queue in time sequence;

[0024] When the blocking queue is full, the write operation is in a blocked waiting state until the blocking queue becomes not full;

[0025] When the blocking queue is empty, the read operation is in a blocked waiting state until the blocking queue becomes non-empty.

[0026] An adaptive target detection system for TOF point cloud, characterized in that the system comprises: a plurality of TOF cameras, a receiving concurrent consumption module deployed on a client, and a point cloud data processing and recognition module deployed on a plurality of distributed servers; wherein,

[0027] The receiving concurrent consumption module is used to receive the TOF point cloud data of the target to be detected collected by the TOF camera, and write them into the corresponding blocking queues respectively, and is also used to read the TOF point cloud data from each blocking queue and input them into the point cloud data processing and recognition module;

[0028] The point cloud data processing and identification module includes: a preliminary denoising unit, a depth range acquisition unit, a Gaussian mixture model, a denoising unit and a target detection model; wherein,

[0029] The preliminary denoising unit is used to sequentially use the DB-SCAN algorithm and the binary iterative k-means algorithm on the TOF point cloud data to obtain the point cloud data after preliminary denoising;

[0030] The depth range acquisition unit is used to use a k-means algorithm to obtain the depth range of the target distribution to be detected by the depth information of the point cloud data after preliminary denoising;

[0031] The Gaussian mixture model is used to cluster the preliminarily denoised point cloud data within the depth range to determine the point cloud data corresponding to different targets to be detected.

[0032] The denoising unit is used to perform local denoising processing on the point cloud data of each target to be detected respectively by using the iterative bilateral filtering algorithm.

[0033] The target detection model is used to complete the recognition of each target according to the denoised point cloud data of each target to be detected.

[0034] As an improvement of the above system, the system further includes an interactive display module deployed on the client side, which is used to display the target detection process and recognition results of the TOF point cloud.

[0035] Compared with the prior art, the advantages of the present invention are as follows:

[0036] 1. Based on the idea of machine learning, the present invention proposes an end-to-end adaptive denoising algorithm, and then completes the practical denoising task for the data directly collected by TOF. Based on the above denoised data, combined with deep learning, the target detection task of three-dimensional point cloud is carried out, greatly improving the recognition accuracy.

[0037] 2. Combining machine learning with traditional filtering algorithms, a practical denoising task can be completed without too many subjective parameters being set artificially. At the same time, deep learning is used to identify the effective data, further realizing the end-to-end target detection goal, overcoming most of the human dependence of only using traditional filtering schemes, and making automatic target detection under large-scale data possible.

[0038] 3. In terms of system architecture, means such as blocking queues, distributed systems, and load balancing are used to ensure the detection efficiency of the camera at high frame rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flowchart of the adaptive target detection method for TOF point cloud of the present invention;

[0040] Figure 2 is a schematic diagram of the system composition architecture of Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] The present invention discloses an adaptive target detection method and system for TOF point clouds, which are generally divided into an algorithm part and an architecture design part. The algorithm part combines machine learning and traditional filtering algorithms, and uses DB-SCAN, PCA, K-means, GMM, bilateral filtering and corresponding improved algorithms to realize denoising and detection of the original point cloud. Finally, the detected targets are recognized through the PointNet network. The architecture part mainly includes a host computer, a server and front-end visualization. The host computer improves the throughput by maintaining a blocking queue, the server is distributedly deployed to provide data processing capabilities adapted to the camera shooting frame rate, and the front-end realizes interactive visualization of point cloud data.

[0042] The technical solution of the present invention will be described in detail below with reference to the drawings and embodiments.

[0043] Embodiment 1

[0044] As Figure 1 shown, Embodiment 1 of the present invention proposes an adaptive target detection method for TOF point clouds, including an adaptive denoising, an extraction algorithm and a target recognition algorithm. The method includes the following steps:

[0045] (1) Using the DB-SCAN algorithm to remove the outlier points of multipath interference;

[0046] (2) Using the binary iterative k-means to further detect and remove the ground noise points;

[0047] (3) Finding the approximate depth range of the target distribution through k-means;

[0048] (4) Using GMM to cluster the point cloud data within this range;

[0049] (5) Using the iterative bilateral filtering algorithm to denoise each target.

[0050] (6) Constructing and training a PointNet deep network model, and inputting the above result targets into the model for recognition respectively.

[0051] The following will be described in detail.

[0052] In step (1), DB-SCAN is a density-based clustering method, which tends to cross the class division boundary through the low-density area. Compared with the previous methods based on statistical probability, this method does not require a clear number of clusters K. It can automatically divide the appropriate number of clusters according to the data distribution characteristics, and can accept cluster distributions of any shape. Most importantly, it is more robust to outlier points. Therefore, this method is suitable for removing outlier points.

[0053] In step (2), the binary iterative k-means is used to further detect and remove the ground noise points. This is not a direct clustering task. In the vertical ground dimension, each time k-means only performs binary classification. The category with a larger data volume is extracted, and then binary clustering is performed on it, and so on. After 3-4 rounds, the target area can be found more accurately. This is different from directly dividing the data into 8 / 16 equal parts or performing a one-time clustering with k = 8 / 16 on the data. The method used in the present invention has an obvious precision advantage compared with the latter two. The binary iterative k-means algorithm is an optimized method for the k-means algorithm in this specific problem. The binary here means k = 2, and the iteration means repeatedly executing k-means with k = 2 multiple times. In each loop, one of the two categories selected for classification is used as the input for the next loop. Using the binary and iterative methods on the basis of k-means is also an innovation point of the present invention.

[0054] In step (3), in the process of finding the approximate depth range of the target distribution through k-means, since the ground point cloud noise has been removed in the above steps, the PCA algorithm can be used to find the best mapping hyperplane to maximize the variance of the data point distribution. This dimension actually represents the depth range of the target. Based on this, a k-means clustering is performed once. The selection of k here is one of the more important hyperparameters of the present invention and has a greater impact on the final extraction effect of the present invention.

[0055] K-means is a simple and efficient unsupervised learning algorithm that can be used for the clustering task of point cloud data. The scientific principle is that each cluster is defined by its centroid or prototype μ, the distance is measured using the Euclidean distance, and each node is assigned to the cluster of the nearest centroid. Then, the data is partitioned by optimizing the minimum value of the distance from the data points to the cluster center.

[0056] PCA (Principal Component Analysis), namely the principal component analysis method, is the most widely used data dimensionality reduction algorithm. The main idea of PCA is to map n-dimensional features onto k dimensions. These k dimensions are new orthogonal features, also known as principal components, which are reconstructed from the original n-dimensional features. The work of PCA is to sequentially find a set of mutually orthogonal coordinate axes in the original space. The selection of the new coordinate axes is closely related to the data itself. Among them, the first new coordinate axis is selected in the direction with the largest variance in the original data. The second new coordinate axis is selected in the plane orthogonal to the first coordinate axis where the variance is the largest. The third axis is the one with the largest variance in the plane orthogonal to the first and second axes. And so on, n such coordinate axes can be obtained. By obtaining the new coordinate axes in this way, we find that most of the variance is contained in the first k coordinate axes, and the variance contained in the subsequent coordinate axes is almost 0. Therefore, we can ignore the remaining coordinate axes and only retain the first k coordinate axes containing most of the variance. In fact, this is equivalent to only retaining the dimensional features containing most of the variance and ignoring the feature dimensions with almost 0 variance, achieving the dimensionality reduction processing of data features.

[0057] In step (4), when using GMM for clustering of the point cloud data within this range, parameter training of GMM is required. Here, the dataset must be trained using the real data in this experimental scenario, so that it can learn the Gaussian mixture distribution in this scenario.

[0058] The GMM Gaussian mixture model is essentially a more general case of k-means. It assumes that the samples are sampled from K clusters, each cluster follows a Gaussian distribution, and a cluster is randomly selected from the K clusters with a certain probability, and sample points are randomly sampled from its distribution to obtain the observed data. Therefore, this algorithm is suitable for processing data with a "non-spherical" distribution.

[0059] In step (5), when using the improved iterative bilateral filtering algorithm, the main optimization point is that based on the traditional filtering algorithm, it is carried out iteratively, and the parameter threshold for each iteration decreases exponentially. While weakening the influence of the initial parameter value on the filtering effect, it avoids filtering out too many valid data points and improves the recognition rate of the subsequent deep neural network.

[0060] In step (6), the trained PointNet neural network is used to detect the categories of each extracted target. Here, PointNet is a multi-classification network, so only the classification network in the original network needs to be retained, and the semantic segmentation network is not used in this invention.

[0061] PointNet is the first deep neural network applicable to point cloud data format. It solves the problem of the disorder of point cloud data and provides a network structure supporting rotational invariance, which is used in this system to complete the target recognition task after denoising.

[0062] The system components include: non-blocking data acquisition at the device end, realizing read-write separation through a blocking queue to improve the shooting frame rate. The data is remotely transmitted to the server through the computer network for algorithm processing. Multiple servers are deployed distributively to handle the data volume collected by the camera. At the same time, load balancing scheduling needs to be carried out for the distributed servers to handle the high-concurrency scenario of data transmission. When necessary, a distributed lock is adopted to solve the synchronization control problem.

[0063] According to Figure 1 the algorithm steps shown below, the following specific solutions are carried out:

[0064] 1. As mentioned before, TOF is vulnerable to multipath reflection, and thus the point cloud data captured in an enclosed space is extremely likely to generate outliers. Through the DB-SCAN algorithm, first find and eliminate the outliers generated thereby.

[0065] 2. On the above basis, since there must be data points generated by ground reflection in the data, and the input of the deep neural network is only the point cloud of the target object, it is necessary to detect and remove this part of the data points. Combining the characteristics of this point cloud data and for the purpose of achieving adaptiveness, here the direction perpendicular to the ground is used as the measurement dimension, and in a binary form, iterative k-means clustering is carried out, that is, the data points are divided into two categories each time, and the category with more data points is selected as the input for the next round of iteration. Finally, iterating 3 - 4 rounds can adaptively find a better segmentation point.

[0066] 3. On the above basis, directly perform k-means clustering on the depth information again to detect the approximate depth range where the target to be measured is located (the best mapping hyperplane can be found through PCA). The specific parameter k here needs to be set manually and is also one of the more important parameters of this algorithm. When the scene changes greatly, the change in the value of k has a greater impact on the final effect of this algorithm, but the value of k does not need to change significantly in the same scene. Therefore, it can meet the demand for large-scale automatic detection in one experimental scene.

[0067] 4. On the above basis, directly use GMM to detect multiple targets for the three-dimensional data. At this time, the postures, sizes, etc. of each target are unknown, so the distribution is no longer balanced. Obviously, the GMM algorithm is more suitable for dealing with the problems in the current scene.

[0068] 5. For each of the above-detected target objects, a traditional filtering algorithm is used for final denoising. However, since most traditional filtering algorithms rely on manually selected parameters, an improvement is made here. For each set of data, iterative filtering is performed on it, and the specific parameters will change exponentially according to the number of iterative rounds, so as to weaken the influence of the initial parameter values on the final processing effect and achieve the purpose of self-adaptation.

[0069] 6. Finally, the above-extracted target point cloud is input into the PointNet deep neural network, which is trained as a multi-classification network. That is, after taking a target point cloud to be detected as input, the category of the target is output, and the specific target extraction task has been completed in the previous process. After the model is pre-trained using the PASCAL3D+ dataset, it is then fine-tuned using the dataset of this scenario to improve the recognition accuracy of the images captured by the TOF camera used in this system.

[0070] Embodiment 2

[0071] As Figure 2 shown, Embodiment 2 of the present invention proposes an adaptive target detection system for TOF point clouds.

[0072] System architecture part: including a host computer, a server, front-end visualization, etc.

[0073] The host computer end provides a non-blocking data saving strategy to improve the camera shooting frame rate.

[0074] The server is deployed distributively, and multiple hosts provide computing resources to improve the computing power.

[0075] The processed data is still saved in the format of point cloud data. At the same time, the front end supports the interactive visualization display of the point cloud data to improve the user experience of the users, facilitate the evaluation of the results and make corresponding parameter adjustments.

[0076] A CCD is connected behind the TOF camera, which is responsible for transmitting the captured data. After the CCD, a microcomputer is connected through a network card. The captured data will be directly transmitted to the microcomputer through the network cable. The host computer written in the microcomputer is responsible for receiving the incoming data and uploading the received data to the server and requesting processing. As Figure 1As shown in the figure, to prevent the camera from entering a blocked state during data transmission, the host computer program will maintain a blocking queue in the data reception section. The program can simply store the received data in the blocking queue and then immediately return. The consumer side will consume the data simultaneously in the form of multiple threads to adapt to the relatively high production frequency of the producer. It will retrieve the captured data from the blocking queue and upload it to the server for processing. Note that any implementation of a blocking queue, including existing message queue middleware and any data structure that meets the definition of a blocking queue implemented through self-coding, should fall within the scope of the claims.

[0077] For a large number of data requests in a high-concurrency scenario, the present invention adopts a distributed deployment method to achieve load balancing. When each server receives a data request, it will call its own algorithm processing module (such as the above algorithm section) and return the result to the front-end for interactive display.

[0078] When the front-end receives the processed point cloud data file, it will perform interactive visualization on this part of the data. People can observe the corresponding three-dimensional point cloud scene by dragging the mouse, using the scroll wheel, etc.

[0079] The host computer side realizes non-blocking data read-write separation through a blocking queue to improve the camera shooting frame rate. In a blocking queue (BlockingQueue), it is a queue that supports two additional operations. These two additional operations are: when the queue is empty, the thread that retrieves elements will block and wait until the queue becomes non-empty. When the queue is full, the thread that stores elements will block and wait for space to become available in the queue. Blocking queues are often used in producer-consumer scenarios. The producer is the thread that adds elements to the queue, and the consumer is the thread that takes elements from the queue. The blocking queue is the container where the producer stores elements, and the consumer only takes elements from the container. Its specific implementation includes but is not limited to self-coding implementation or using existing various message middleware such as Kafka, etc.

[0080] The captured data is sent to the algorithm processing server through a certain data transmission medium. This data transmission medium includes but is not limited to the Internet, local area network, wired data transmission, etc.

[0081] Distributed deployment servers. Multiple hosts have the same algorithm module. The client does not directly request each server but accesses the load balancing server, which forwards the request to a specific server according to the load balancing algorithm, so that each server bears a comparable request pressure and avoids crashing.

[0082] The front-end uses a certain client program (including application client programs, web browser pages, and embedded application programs all fall within the scope of the claims) to achieve the interactive visualization display of three-dimensional point cloud data.

[0083] The specific technical solution is as follows: The system includes: several TOF cameras, a receiving and concurrent consumption module deployed on the client side, and a point cloud data processing and recognition module deployed on multiple distributed servers; among them,

[0084] The receiving and concurrent consumption module is used to receive the TOF point cloud data of the target to be detected collected by the TOF camera, write them into the corresponding blocking queues respectively, and is also used to read the TOF point cloud data from each blocking queue and input it into the point cloud data processing and recognition module;

[0085] The point cloud data processing and recognition module includes: a preliminary denoising unit, a depth range acquisition unit, a Gaussian mixture model, a denoising unit, and a target detection model; among them,

[0086] The preliminary denoising unit is used to use the DB-SCAN algorithm and the binary iterative k-means algorithm on the TOF point cloud data in sequence to obtain the preliminarily denoised point cloud data;

[0087] The depth range acquisition unit is used to use the k-means algorithm on the depth information of the preliminarily denoised point cloud data to obtain the depth range where the target to be detected is distributed;

[0088] The Gaussian mixture model is used to cluster the preliminarily denoised point cloud data within the depth range to determine the point cloud data corresponding to different targets to be detected;

[0089] The denoising unit is used to use the iterative bilateral filtering algorithm to perform local denoising processing on the point cloud data of each target to be detected respectively;

[0090] The target detection model is used to complete the recognition of each target according to the denoised point cloud data of each target to be detected.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not restrictive. Although the present invention has been described in detail with reference to the embodiments, those of ordinary skill in the art should understand that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. An adaptive target detection method for TOF point clouds, the method comprises: Receiving TOF point cloud data of the target to be detected collected by a TOF camera; Successively using the DB-SCAN algorithm and the binary iterative k-means algorithm on the TOF point cloud data to obtain preliminarily denoised point cloud data; Using the k-means algorithm for the depth information of the preliminarily denoised point cloud data to obtain the depth range where the target to be detected is distributed; Using a pre-established and trained Gaussian mixture model to cluster the preliminarily denoised point cloud data within the depth range to determine the point cloud data corresponding to different targets to be detected; Using the iterative bilateral filtering algorithm to perform local denoising processing on the point cloud data of each target to be detected respectively; Respectively inputting the denoised point cloud data of each target to be detected into a pre-established and trained target detection model to complete the recognition of each target.

2. The adaptive target detection method for TOF point clouds according to claim 1, wherein, By using the DB-SCAN algorithm on the TOF point cloud data, the outlier points generated due to the influence of multipath reflection are found and eliminated.

3. The adaptive target detection method for TOF point clouds according to claim 1, wherein, Using the binary iterative k-means algorithm for the point cloud data after eliminating the outlier points, specifically including: Taking the direction perpendicular to the ground as the measurement dimension, using the k-means algorithm to perform binary clustering on the point cloud data after eliminating the outlier points to obtain two types of point cloud data; Selecting the type of point cloud data with more data points from them, and using the k-means algorithm again for binary clustering. Repeating this step until the set number of rounds is reached to obtain the segmentation points that meet the preset requirements, finding the ground noise points to be eliminated, and eliminating them.

4. The adaptive target detection method for TOF point clouds according to claim 1, wherein, Using the k-means algorithm for the depth information of the preliminarily denoised point cloud data to obtain the depth range where the target to be detected is distributed; specifically including: Using the PCA algorithm for the point cloud data within each target to be detected area to obtain the best mapping hyperplane, so that the variance of the distribution of the point cloud data within the target to be detected area is the largest, performing dimensionality reduction mapping on the point cloud data, and recording the corresponding relationship between the data points before and after the dimensionality reduction mapping; According to the preset number of classifications k, performing one-time k-means clustering on the dimensionality-reduced point cloud data to determine the data points of the target to be detected, and according to the data relationship before and after the dimensionality reduction mapping, finding the corresponding original point cloud data, retaining this part of the data, eliminating other data points, and combining the depth information in the point cloud data to obtain the depth range where the target to be detected is distributed.

5. The adaptive target detection method for TOF point clouds according to claim 1, wherein, The target detection model is a PointNet neural network, the input is the denoised point cloud data of the target to be measured, and the output is the target recognition result.

6. The adaptive target detection method for TOF point clouds according to claim 1, wherein, Before receiving the TOF point cloud data of the target to be detected collected by the TOF camera, it further includes: establishing a blocking queue for each TOF camera in advance to cache the TOF point cloud data; the blocking queue supports first-in-first-out read and write operations, specifically including: Through the write operation, the TOF point cloud data of the target to be detected is stored in the blocking queue according to the time sequence, and through the read operation, the TOF point cloud data is obtained from the blocking queue according to the time sequence; When the blocking queue is full, the write operation is in a blocking waiting state until the blocking queue becomes non-full; When the blocking queue is empty, the read operation is in a blocking waiting state until the blocking queue becomes non-empty.

7. An adaptive target detection system for TOF point clouds, characterized in that, The system includes: a number of TOF cameras, a receiving and concurrent consumption module deployed on the client, and a point cloud data processing and recognition module deployed on multiple distributed servers; among them, The receiving and concurrent consumption module is used to receive the TOF point cloud data of the target to be detected collected by the TOF camera, and write it into the corresponding blocking queue respectively, and is also used to read the TOF point cloud data from each blocking queue and input it into the point cloud data processing and recognition module; The point cloud data processing and recognition module includes: a preliminary denoising unit, a depth range acquisition unit, a Gaussian mixture model, a denoising unit, and a target detection model; among them, The preliminary denoising unit is used to sequentially use the DB-SCAN algorithm and the binary iterative k-means algorithm on the TOF point cloud data to obtain the preliminarily denoised point cloud data; The depth range acquisition unit is used to use the k-means algorithm for the depth information of the preliminarily denoised point cloud data to obtain the depth range in which the target to be detected is distributed; The Gaussian mixture model is used to cluster the preliminarily denoised point cloud data within the depth range to determine the point cloud data corresponding to different targets to be detected; The denoising unit is used to use the iterative bilateral filtering algorithm to perform local denoising processing on the point cloud data of each target to be detected respectively; The target detection model is used to complete the recognition of each target according to the denoised point cloud data of each target to be detected.

8. The adaptive target detection system for TOF point clouds according to claim 7, characterized in that, The system further includes an interactive display module deployed on the client, which is used to display the target detection process and recognition results of the TOF point cloud.

Images