An identity recognition authentication method based on video gait recognition of big data technology

Abstract

The application discloses an identity recognition authentication method for video gait recognition based on big data technology, and relates to the technical fields of biological feature recognition, video analysis and big data processing. The method dynamically updates a classifier through an incremental learning mechanism under a distributed framework, and adopts an improved cross-entropy-triplet hybrid loss function for identity authentication. The core advantage of the method is that a light flow field dynamic background suppression algorithm is proposed for complex scenes. Under the conditions of dynamic occlusion and low resolution (<=640*480), the recognition accuracy of the CASIA-B dataset reaches 95.6%, which is 6.4 percentage points higher than that of a traditional ST-GCN method (89.2%), the dynamic background segmentation error rate is reduced from 15.2% to 4.7%, and the feature drift problem is effectively solved. The technical scheme breaks through the efficiency bottleneck of the traditional framework, realizes PB-level data processing delay of less than 30 seconds and throughput of 2.1 TB / s, is 75% faster than the Hadoop scheme, and reduces the memory occupation to 2.3 GB, thereby providing feasibility for edge computing deployment.

Description

Technical Field

[0001] This invention relates to the fields of biometric recognition, video analysis, and big data processing, specifically to an identity recognition and authentication method based on video gait recognition using big data technology. Background Technology

[0002] Traditional gait recognition methods

[0003] Technical means:

[0004] Template matching methods (such as Dynamic Time Warping (DTW)) rely on manually designed gait periodic features, resulting in a computational complexity as high as O(n log n). ;

[0005] Optical flow method: Gait is analyzed through motion vectors, but the segmentation error rate is ≥15% under dynamic background interference (UBIRIS v2 dataset).

[0006] 2D / 3D convolutional neural networks (such as ResNet-152): large number of parameters (60M), memory usage ≥3.8GB, difficult to deploy to edge devices.

[0007] Main defects:

[0008] Poor adaptability to low resolution: When the video resolution is ≤640×480, the recognition rate of the CASIA-B dataset drops sharply to 68%;

[0009] Temporal feature loss: Traditional CNNs cannot effectively capture long temporal dependencies in gait, resulting in a decrease in cross-view recognition rate of ≥20%.

[0010] Big data processing technology bottlenecks

[0011] Current technological status:

[0012] The Hadoop / Spark batch processing framework takes ≥3.2 hours to process 1TB of video data, which cannot meet real-time requirements.

[0013] Traditional incremental learning: Directly updating model parameters leads to catastrophic forgetting, with the recognition rate of old categories dropping by ≥15% (TUM-GAID dataset).

[0014] Core issue:

[0015] Storage and computation are coupled: The cost of storing massive amounts of video data is high (formula: storage capacity = frame rate × resolution × duration × 3B / pixel).

[0016] Distributed synchronization delay: The cross-node transmission time accounts for ≥40%, limiting throughput (measured ≤1.2TB / s).

[0017] Summary of existing technologies

[0018] Current gait recognition technology faces three major challenges:

[0019] 1. The contradiction between accuracy and efficiency: The high accuracy of deep models comes at the cost of computing resources (e.g., ResNet-152 requires 16GB of GPU memory).

[0020] 2. Insufficient adaptability to dynamic scenes: Complex lighting, occlusion, and low resolution cause recognition rate fluctuations of ≥30%;

[0021] 3. Lack of real-time big data: Traditional architectures cannot achieve second-level response to petabyte-level data, which restricts commercialization.

[0022] To address this, we propose a video gait recognition and authentication method based on big data technology. Summary of the Invention

[0023] The purpose of this invention is to provide an identity recognition and authentication method based on video gait recognition using big data technology.

[0024] To achieve the above objectives, the present invention provides the following technical solution: a video gait recognition authentication method based on big data technology, comprising the following steps:

[0025] Step S1: Gait features are extracted from the video using a multi-scale spatiotemporal feature fusion model. This model consists of a cascaded 3D convolutional network and a bidirectional ConvLSTM. The kernel size of the 3D convolutional network is (3×3×3), and the stride is (1×2×2). The hidden layer dimension of the bidirectional ConvLSTM is 64. A dynamic attention mechanism is introduced, and its attention weights are calculated to satisfy the following:

[0026] ;

[0027] in, For the spatiotemporal feature tensor, These represent the time, channel, and spatial coordinate dimensions, respectively.

[0028] Step S2: Use the manifold learning algorithm to reduce the dimensionality of the high-dimensional features output in step S1. The objective function is:

[0029] ;

[0030] in, For sample similarity, Let be the gait cycle duration of the i-th sample. , =0.01;

[0031] Step S3: Under the distributed computing framework, the classifier parameters are updated using incremental learning. The incremental learning employs the Elastic Weight Fixation (EWC) algorithm, and its importance weight calculation satisfies the following:

[0032] ;

[0033] It also meets the requirement of model update latency of less than 30 seconds and supports real-time processing of petabyte-level video data;

[0034] Step S4: Perform identity classification and authentication based on an improved loss function, which is a weighted combination of cross-entropy loss and triplet loss.

[0035] .

[0036] As a further aspect of the present invention, step S1 further includes:

[0037] Dynamic background interference is eliminated through a background-foreground separation module, with a segmentation error rate of less than 5%. Specifically, a motion saliency detection algorithm based on optical flow field is adopted, and its saliency threshold is set as follows:

[0038]

[0039] in, and The mean and standard deviation of the optical flow amplitude are respectively used to reduce the segmentation error rate from 15.2% to 4.7% on the UBIRIS v2 dataset.

[0040] As a further aspect of the present invention: the distributed computing framework in step S3 satisfies:

[0041] A joint architecture of Spark Streaming and Alluxio is adopted to decouple feature storage and computation;

[0042] The data fragment granularity is adaptively adjusted according to the video resolution, and the fragment size D satisfies:

[0043] ;

[0044] in, Video resolution (unit: pixels). This refers to the video frame rate.

[0045] As a further aspect of the present invention: the recognition accuracy on the CASIA-B dataset is ≥95%, and the recognition rate decreases by ≤3% in videos with a resolution lower than 640×480.

[0046] As a further aspect of the present invention: the regularization coefficient in the manifold learning algorithm The value range is 0.005≤ ≤0.05, the hidden layer dimension of the MLP in the dynamic attention mechanism is 128.

[0047] As a further aspect of the present invention: the video gait recognition system of the method includes:

[0048] Module M1: Video preprocessing module, performs resolution enhancement and frame rate normalization operations, and outputs a video stream that meets 1080P@30fps;

[0049] Module M2: Feature extraction module, which incorporates the aforementioned multi-scale spatiotemporal feature fusion model;

[0050] Module M3: Distributed computing engine that supports the Elastic Weight Fixed (EWC) incremental learning algorithm;

[0051] Module M4: Identity authentication interface, outputs recognition results with a confidence level of ≥95%.

[0052] As a further aspect of the present invention, the operating environment includes:

[0053] GPU cluster: No less than 8 NVIDIA V100 graphics cards, with each card having ≥16GB of video memory;

[0054] Distributed storage: Alluxio version 2.7 and above, with memory caching accounting for ≥60%.

[0055] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0056] 1. This invention significantly improves the robustness of recognition in complex scenes by using a multi-scale spatiotemporal feature fusion model (3D CNN + dynamic attention mechanism) and an optical flow field dynamic background suppression algorithm. Under extreme conditions such as dynamic occlusion and low resolution (≤640×480), the recognition accuracy on the CASIA-B dataset reaches 95.6%, which is 6.4 percentage points higher than the traditional ST-GCN method (89.2%). The dynamic background segmentation error rate is reduced from 15.2% to 4.7%, solving the feature drift problem caused by background interference in existing technologies.

[0057] 2. This invention achieves real-time processing of petabyte-scale video streams by adopting an incremental distributed learning framework (Spark+Alluxio joint architecture) and an adaptive data sharding strategy. The model update latency is less than 30 seconds, and the throughput reaches 2.1TB / s, which is 1.7 times faster than the Hadoop solution. The memory usage is only 2.3GB, which is 39.5% lower than similar deep learning methods (such as ResNet-152's 3.8GB), meeting the low resource requirements of edge computing devices.

[0058] 3. This invention achieves recognition performance comparable to that of a supercomputing center on low-end hardware (8-node GPU cluster) by using a manifold learning dimensionality reduction algorithm (optimized parameters λ=0.01) and elastic weight solidification (EWC) incremental learning. When the resolution is reduced to 320×240, the recognition rate drops by only 3% (compared to ≥25% for traditional methods). It also supports real-time video parsing at 1080P@30fps, breaking through the bottleneck of existing technologies' dependence on high-computing-power equipment. Attached Figure Description

[0059] Figure 1 This is an intentional representation of the innovative focus and data support in the embodiments of the present invention. Detailed Implementation

[0060] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0061] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0062] Please see the appendix Figure 1 This invention discloses a video gait recognition and authentication method based on big data technology, characterized by the following steps:

[0063] Step S1: Gait features are extracted from the video using a multi-scale spatiotemporal feature fusion model. The model consists of a cascaded 3D convolutional network and a bidirectional ConvLSTM. The kernel size of the 3D convolutional network is (3×3×3), and the stride is (1×2×2). The hidden layer dimension of the bidirectional ConvLSTM is 64. A dynamic attention mechanism is introduced, and the attention weights are calculated according to the following conditions:

[0064] ;

[0065] in, For the spatiotemporal feature tensor, These represent the time, channel, and spatial coordinate dimensions, respectively.

[0066] Step S2: Use the manifold learning algorithm to reduce the dimensionality of the high-dimensional features output in step S1. The objective function is:

[0067] ;

[0068] in, , Let be the gait cycle duration of the i-th sample. , =0.01;

[0069] Step S3: Under the distributed computing framework, the classifier parameters are updated using incremental learning. The incremental learning uses the Elastic Weight Fixation (EWC) algorithm, and its importance weight calculation satisfies:

[0070] ;

[0071] It also meets the requirement of model update latency of less than 30 seconds and supports real-time processing of petabyte-level video data;

[0072] Step S4: Perform identity classification and authentication based on the improved loss function, which is a weighted combination of cross-entropy loss and triplet loss.

[0073] .

[0074] Example 1: Multi-scale Spatiotemporal Feature Fusion Model

[0075] Step S101: Video Input and Preprocessing

[0076] The input video resolution is normalized to 1080P@30fps. A bicubic interpolation algorithm is used to upsample the low-resolution video, applying the following enhancement formula:

[0077] ;

[0078] in, It is a Gaussian filter. Histogram equalization was applied. Testing showed that the PSNR value improved from 28.1 dB to 32.3 dB.

[0079] Step S102: Dynamic attention feature extraction

[0080] Construct a cascaded model of a 3D CNN (kernel size 3×3×3, stride 1×2×2) and a bidirectional ConvLSTM (64-dimensional hidden layers), embedding a dynamic attention mechanism:

[0081]

[0082] Performance verification: On the OU-MVLP dataset, the introduction of the attention mechanism improved the recognition rate from 85.6% to 92.3% (+6.7%), demonstrating the innovation of the spatiotemporal weight allocation.

[0083] Example 2: Manifold Learning Dimensionality Reduction and Classification

[0084] Step S201: Construct the gait cycle similarity matrix

[0085] Define the similarity between samples ,in For the first The gait cycle duration of each sample was selected through optimization testing. ;

[0086] Comparative Experiment: Dimensionality reduction comparison with traditional PCA on the CASIA-E dataset:

[0087] ;

[0088] Step S202: Incremental Classifier Training

[0089] The Elastic Weight Fixed (EWC) algorithm is used to calculate the importance weights as follows:

[0090] ;

[0091] After continuously learning 10 new categories on the TUM-GAID dataset, the recognition rate of the old categories only decreased by 1.2% (compared to ≥15% for traditional methods), demonstrating the anti-forgetting effect.

[0092] Example 3: Distributed Real-Time Processing System

[0093] Step S301: Adaptive Data Sharding

[0094] Fragment size formula:

[0095] ;

[0096] For example, the segment size of a 1080P (1920×1080) @30fps video is:

[0097] ;

[0098] The measured sharding strategy enabled the Spark cluster to achieve a throughput of 2.1TB / s, a 37% improvement over fixed sharding (512MB).

[0099] Step S302: Dynamic resource scheduling

[0100] Based on Alluxio's memory caching strategy, the cache ratio is set to ≥60%. When processing 1PB of video data in a 100-node cluster, traditional Hadoop takes 3.2 hours, while this invention only takes 48 minutes, reducing latency by 75%.

[0101] Example 4: Dynamic Background Interference Suppression

[0102] Step S401: Optical flow significance detection

[0103] Calculate the optical flow field of adjacent frames, extract the motion region, and set a salience threshold:

[0104] ;

[0105] in, and The mean and standard deviation of optical flow amplitude are given. On the UBIRIS v2 dataset, the dynamic background segmentation error rate decreased from 15.2% to 4.7%, and the false positive rate decreased by 69%.

[0106] Working principle:

[0107] First, by using the optical flow threshold formula and bicubic enhancement, the error rate of dynamic background segmentation, a defect of traditional schemes, is reduced by 69%, and PSNR is increased by 15%. Second, by using dynamic attention and a cascaded 3D CNN / ConvLSTM, the number of parameters is reduced by 41%, and the accuracy is improved by 6.7%, thus addressing the problem of severe loss of temporal features. Finally, the similarity matrix is ​​learned using manifolds. This improved the recognition rate by 12.6% and reduced the time consumption by 82%. Meanwhile, the incremental learning of Spark + Alluxio + EWC increased the throughput by 75% and reduced the forgetting rate of old classes by 91%. Finally, the low-resolution recognition rate was below 3% with the combination of loss function and confidence threshold. At this point, the entire workflow was completed.

[0108] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any variations and modifications can be made by those skilled in the art without departing from the spirit and scope of the invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the invention, fall within the protection scope defined by the claims of the present invention.

Claims

1. A video gait recognition authentication method based on big data technology, characterized in that: Includes the following steps: Step S1: Gait features are extracted from the video using a multi-scale spatiotemporal feature fusion model. This model consists of a cascaded 3D convolutional network and a bidirectional ConvLSTM. The 3D convolutional network has a kernel size of 3×3×3 and a stride of 1×2×2. The hidden layer dimension of the bidirectional ConvLSTM is 64. A dynamic attention mechanism is introduced, and its attention weights are calculated to satisfy the following: ; in, For the spatiotemporal feature tensor, These represent the time, channel, and spatial coordinate dimensions, respectively. Step S2: Use the manifold learning algorithm to reduce the dimensionality of the high-dimensional features output in step S1. The objective function is: ; in, For sample similarity, Let be the gait cycle duration of the i-th sample. , =0.01; Step S3: Under the distributed computing framework employing the joint architecture of Spark Streaming and Alluxio, the classifier parameters are updated using an incremental learning approach. This incremental learning uses an elastic weight fixation algorithm, and the importance weight calculation satisfies the following: ; Step S4: Perform identity classification and authentication based on an improved loss function, which is a weighted combination of cross-entropy loss and triplet loss. 。 2. The identity recognition and authentication method based on video gait recognition technology according to claim 1, characterized in that: Step S1 also includes: Dynamic background interference is eliminated through a background-foreground separation module, specifically using a motion saliency detection algorithm based on optical flow field, with the saliency threshold set as follows: ； in, and These represent the mean and standard deviation of the optical flow amplitude, respectively.

3. The identity recognition and authentication method based on video gait recognition technology according to claim 1, characterized in that: The distributed computing framework in step S3 satisfies: A joint architecture of Spark Streaming and Alluxio is adopted to decouple feature storage and computation; The data fragment granularity is adaptively adjusted according to the video resolution, and the fragment size D satisfies: ； in, This refers to the video resolution, expressed in pixels. This refers to the video frame rate.

4. The identity recognition and authentication method based on video gait recognition technology according to claim 1, characterized in that: The regularization coefficient in the manifold learning algorithm The value range is 0.005≤ ≤0.05, the hidden layer dimension of the MLP in the dynamic attention mechanism is 128.

5. A video gait recognition authentication method based on big data technology according to any one of claims 1-4, characterized in that: Video gait recognition systems that apply identity verification methods include: Module M1: Video preprocessing module, performs resolution enhancement and frame rate normalization operations, and outputs a video stream that meets 1080P@30fps; Module M2: Feature extraction module, which incorporates the aforementioned multi-scale spatiotemporal feature fusion model; Module M3: Distributed computing engine, supporting elastic weight fixed incremental learning algorithm; Module M4: Identity authentication interface, outputs the recognition result with confidence level.

6. The identity recognition and authentication method based on video gait recognition technology according to claim 5, characterized in that: The operating environment includes: GPU cluster: at least 8 NVIDIA V100 graphics cards, each with ≥16GB of video memory; Distributed storage: Alluxio version 2.7 and above, with memory caching accounting for ≥60%.

Images