Video crowd counting method based on cascaded cross-domain feature interaction network

Abstract

The invention discloses a video crowd counting method based on a cascaded cross-domain feature interaction network. The method comprises the following steps: carrying out data enhancement processing of random cutting and horizontal flipping on a current frame and front and back frames of the current frame; and constructing a cross-domain feature interaction network composed of a spatial domain branch and a frequency domain branch. The frequency domain branch extracts frequency domain feature output of different stages through a high and low frequency signal aggregation module and a feature encoder based on adjacent frames; the spatial domain branch is based on a single-frame image, and static spatial semantic features are extracted through a feature encoder. Cascade fusion is carried out on the double-branch features on multiple scales, two-way channel cross attention is utilized to reconstruct time sequence correlation frequency domain features of a current frame, and fusion and reconstruction of the two domain features are achieved through a cross-domain feature mutual modulation module. And after the reconstructed double-branch features are processed by the fusion network, outputting a crowd density map of the current frame by a density regression head. And after training is completed, storing the optimal model for video crowd counting. According to the invention, through cross-domain feature cascade and bidirectional time sequence modeling, the accuracy and robustness of crowd counting in a video scene are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of crowd counting technology, and in particular to a video crowd counting method based on a cascaded cross-domain feature interaction network. Background Technology

[0002] With the rapid development of intelligent surveillance systems and the increasing frequency of large-scale public events, video crowd counting has become a key technology in areas such as public safety, traffic management, and crowd control. The goal of video crowd counting is to accurately estimate the number of people in each video frame. Current mainstream methods address this problem by estimating a crowd density map and summing the density maps.

[0003] In existing research, video people counting can be roughly divided into two categories: spatiotemporal correlation-based methods and multimodal feature fusion-based methods. However, there are two main problems: First, spatiotemporal correlation-based methods mainly capture local spatiotemporal motion in a single spatiotemporal domain, and easily ignore the rich information in other domains. Second, multimodal feature fusion-based methods are mostly derived directly from the motion features between adjacent frames, so they have a strong coupling relationship with the original RGB features. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a video crowd counting method based on a cascaded cross-domain feature interaction network, which utilizes the full interaction between the spatial domain and the frequency domain to accurately estimate the number of people contained in a video frame.

[0005] Technical solution: A video crowd counting method based on cascaded cross-domain feature interaction networks, comprising the following steps:

[0006] S1 divides all frames of the video into training, validation, and test sets according to timeline and a certain ratio, ensuring that the video frames in each set are continuous. Data augmentation is achieved by applying uniform random cropping and horizontal flipping operations to each current frame and the two frames before and after it in the training set, thus obtaining an expanded training sample set. Missing frames at the beginning and end are directly filled in by copying the first and last frames.

[0007] S2, a dual-branch feature coding network is constructed, comprising a frequency domain branch and a spatial domain branch. The frequency domain branch extracts the frequency domain features of each frame based on the image information of the current frame and its preceding and following frames through a high- and low-frequency signal aggregation module, and uses a feature encoder to obtain the multi-stage frequency domain feature representation of each frame. The spatial domain branch takes the image of the current frame as input and extracts spatial static features through a feature encoder. The frequency domain branch and the spatial domain branch are in parallel structure, providing feature input for subsequent cross-domain feature fusion.

[0008] S3. A cascaded cross-domain feature interaction decoding network is constructed. This network first uses the current frame as the center, employing frequency domain features from previous and subsequent frames as auxiliary inputs, and extracts the temporally correlated frequency domain features of the current frame through bidirectional channel cross-attention. Subsequently, the extracted temporally correlated frequency domain features and the spatial domain static features of the current frame are input into the cross-domain feature intermodulation module to achieve mutual fusion and enhancement of the two domain features. The above temporally correlated frequency domain feature extraction and cross-domain feature intermodulation processes are repeatedly executed in a cascaded manner at different stages of the encoded features to gradually strengthen the semantic representation of the temporally correlated frequency domain and spatial domain of the current frame. Finally, the enhanced dual-branch features are concatenated, and channel compression and feature integration are performed through a multi-layer convolutional fusion network to output the crowd density map of the current frame. The estimated number of people in the current frame is obtained through global summation.

[0009] S4. After multiple rounds of training on a cascaded cross-domain feature interaction network that simultaneously includes encoding and decoding, the network is tested on a validation set. The network weights corresponding to the best results on the validation set are saved to obtain the optimal network model. Each current frame and the two frames before and after it are used as samples in the test set and input into the optimal network model to estimate the number of people in each current frame.

[0010] Furthermore, in step S2, the feature encoder uses a ConvNeXt-T model with the global average pooling layer and fully connected layer removed to extract frequency domain features and spatial features of the input frame. The encoder used in this invention can also be replaced by a feature encoder such as ResNet.

[0011] Furthermore, the implementation process of the high- and low-frequency signal aggregation module is as follows: Taking the current frame as an example, the RGB image features of the current frame are first decomposed into a low-frequency sub-band LL and three high-frequency sub-bands LH, HL, and HH using Haar wavelet transform. The low-frequency sub-band LL is used to capture local grayscale changes, while the high-frequency sub-bands are used to capture global structural information. Specifically, LH and HL in the high-frequency sub-bands focus on extracting edge features in the horizontal and vertical directions, respectively, while HH focuses on capturing diagonal texture information. Subsequently, the three high-frequency sub-bands are convolved with 1×1 and then concatenated to obtain the aggregated high-frequency feature representation F. HF :

[0012] F HF =Concat(Conv1×1(LH),Conv1×1(HL),Conv1×1(HH))

[0013] Since the high-frequency components are sparse and mainly consist of local edge and texture features, we first apply F... HF 2×2 max pooling is performed to highlight significant responses, followed by 3×3 depthwise convolution and 2×2 deconvolution to obtain intermediate features F. HF’Then, a spatially gated graph M is generated using the Sigmoid activation function:

[0014] F HF’ =Upsampling(DWConv(MaxPool(F HF )))

[0015] M = Sigmoid(F) HF’ )

[0016] Next, it is multiplied element-wise with the low-frequency component to enhance its characterization, and then the fused features are obtained through residual connection:

[0017] L out =M⊙LL+LL

[0018] Similarly, the same high and low frequency signal aggregation module is also applied to the previous and next frames, and then frequency domain feature coding is performed through the backbone network.

[0019] 5. Further, in step S3, the bidirectional channel cross-attention module uses adjacent preceding and following frames as aids to construct forward and backward channel attention respectively, thereby extracting the temporally correlated frequency domain features of the current frame, and then weightedly fuses the two branch features. Channel attention is used to enhance salient channels and suppress irrelevant channels. Taking the construction of forward channel attention as an example, the calculation expressions for query Q1, key K, and value V in channel attention are as follows:

[0020] K = Re(W) k ·GAP(C4))

[0021] V = Re(W) v ·GAP(C4))

[0022] Q1 = W q1 ·GAP (P4)

[0023] Among them, W k W v and W q1 These are all learnable parameter matrices used to enhance the model's fitting ability; GAP() represents global average pooling; Re(·) represents the reshape operation. C4 and P4 represent the frequency domain features of the current frame and the previous frame, respectively, output by the ConvNeXt model in stage 4.

[0024] In the forward channel attention, the query Q1 of the previous frame is multiplied by the key K of the current frame to obtain the attention matrix A1. Then, A1 is multiplied by the value vector V and activated by the Sigmoid activation function to obtain the forward channel weight α1. The calculation expressions for A1 and α1 are as follows:

[0025]

[0026] in, ∑ represents matrix multiplication; Sigmoid() and Softmax() represent activation functions; ⊙ represents element-wise multiplication; C is the number of channels; A1 reflects the correlation between the channels of the previous frame and the current frame, ∑ dim =1A1 indicates summation along the first dimension;

[0027] Similarly, for constructing the backward channel attention, the backward channel weight α2 guided by the next frame can be obtained only through C4 and the frequency domain feature N4 of the next frame output from the fourth stage. Then, the weights of the two directions are fused using a weighted average:

[0028] α=λ1α1+λ2α2

[0029] s.tλ1+λ2=1

[0030] Where λ1 and λ2 are learnable parameters;

[0031] Finally, the fused weights α are reshaped into R. C×1×1 Then multiply it with the current frame feature C4 to obtain the temporal reconstruction feature C4′:

[0032] C4′=α⊙C4

[0033] Through this bidirectional modeling mechanism, the model can adaptively highlight important channels and suppress redundant channels under the guidance of the temporal context, so that the features of the current frame contain the temporal change information of the previous and next frames, thereby improving the representation ability and stability of the features of the current frame.

[0034] 6. Further, in step S3, the cross-domain feature mutual modulation module refers to the mutual modulation and enhancement of the spatial domain features of the current frame and the frequency domain features reconstructed based on temporal correlation. One feature can be used as the source domain and the other as the target domain. A spatial weight map is generated using the source domain features to modulate the target domain features. The source domain and target domain can be interchanged. Specifically, assuming the spatial domain features of the current frame are used as the source domain F... src Using frequency domain features reconstructed based on temporal correlation as the target domain F tar First, salient features are extracted from the source domain features by concatenating the results of max pooling and average pooling, followed by convolution.

[0035]

[0036] Among them, MaxPool() represents max pooling; AvgPool() represents average pooling; Concat() represents concatenation operation; Conv 1×1 Represents a 1×1 convolution;

[0037] Max pooling extracts the most salient response at each spatial location, while average pooling extracts the average response; combining the two provides a more comprehensive feature representation. Then, salient features are further processed... After performing a linear transformation and passing it through the Sigmoid activation function, the source domain spatial attention map S is obtained. att :

[0038]

[0039] Where w and b are learnable parameters;

[0040] Then, the source domain spatial attention map S is used. att With target feature F tar Element-wise multiplication is performed, and residual connections are added to preserve the original information and avoid network degradation, resulting in frequency domain features modulated with spatial domain features:

[0041] F tar ′=S att ⊙F tar +F tar

[0042] Similarly, the frequency domain features reconstructed based on temporal correlation are used as the source domain F. src Using the spatial domain features of the current frame as the target domain F tar Then, the spatial domain features can be modulated and enhanced using the frequency domain features of the current frame.

[0043] 7. Further, in step S2, the specific structure of the fusion network is as follows:

[0044] C(256,3,1)-GC(128,3,1)-GC(64,3,1)

[0045] Where C(i,j,k) represents a convolutional layer with i convolutional kernels, a kernel size of j×j, and an inflation rate of k, and G is the GeLU activation function.

[0046] Compared with the prior art, the significant advantages of this invention are as follows:

[0047] 1. This invention proposes a cascaded cross-domain feature interaction network, which fully explores the complementary information between the spatial domain and the frequency domain, effectively realizing the fusion of the two features. It exhibits better performance, stronger generalization ability and higher robustness in video crowd counting tasks. Unlike traditional models that only rely on the RGB domain or motion domain, this invention starts from a cross-domain perspective and constructs a unified framework that can simultaneously model the relationship between the spatial and frequency domains, thereby significantly improving the accuracy of the model.

[0048] 2. This invention proposes a high- and low-frequency signal aggregation method. Addressing the problem of information fragmentation when processing high- and low-frequency features separately, it constructs a unified frequency domain feature representation. By fully utilizing the complementarity between high-frequency details and low-frequency structure, this method effectively enhances the integrity and robustness of frequency domain modeling.

[0049] 3. This invention designs a bidirectional channel cross-attention module that can inject temporal context information into the frequency domain representation, enabling adaptive feature calibration between adjacent frames, thereby improving temporal modeling capabilities. This module achieves cross-frame channel interaction at different scales, effectively focusing on key features and suppressing redundant features, while maintaining linear computational complexity and significantly enhancing the model's temporal consistency and density prediction accuracy.

[0050] 4. This invention proposes a cross-domain feature intermodulation module to enhance the expressive power of frequency and spatial domain features through cross-domain interaction. This module establishes a bidirectional modulation mechanism between the frequency and spatial features of the current frame, enabling the two types of features to guide each other. A spatial weight map is generated from the source domain features and used to modulate the target domain features, thereby achieving adaptive fusion of cross-domain information. The source and target domains are flexibly interchangeable, achieving information fusion at multiple semantic levels through cross-scale interaction. This allows both spatial and frequency domain features to be enhanced in different directions, significantly improving the feature representation capability. Attached Figure Description

[0051] Figure 1 This is the overall flowchart of the present invention;

[0052] Figure 2 This is a diagram of the cascaded cross-domain feature interaction network framework of the present invention;

[0053] Figure 3 This is a framework diagram of the high and low frequency signal aggregation module of the present invention;

[0054] Figure 4 This is a framework diagram of the bidirectional channel cross-attention module of the present invention;

[0055] Figure 5 This is a framework diagram of the cross-domain feature mutual modulation module of the present invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0057] like Figure 1 The diagram shown is a flowchart of a video audience counting method based on a cascaded cross-domain feature interaction network provided by the present invention. Figure 2 The following is a diagram of the cascaded cross-domain feature interaction network framework. The implementation steps of this invention are as follows:

[0058] Step 1: Data preprocessing;

[0059] Specifically, the steps include the following:

[0060] The last 10% of the video frames in the training set were selected as the validation set. Before feature extraction, random cropping of the same region was performed on every three consecutive video frames, and the cropped region was uniformly horizontally flipped with a probability of 0.5 to achieve data augmentation.

[0061] Step two involves constructing a cascaded cross-domain feature interaction network. This includes the following steps:

[0062] Step 21 involves obtaining high-frequency and low-frequency signals by performing wavelet transform on three consecutive video frames, and then effectively fusing the high-frequency texture features and low-frequency global information. Specifically, this includes the following steps:

[0063] For consecutive RGB video frames I∈R T×3×H×W Where 3, H, and W represent the number of channels, height, and width of the video frame, respectively, and the following operations are performed simultaneously:

[0064] like Figure 3 As shown, for an RGB video frame, the current frame is first decomposed into a low-frequency sub-band LL and three high-frequency sub-bands LH, HL, and HH using Haar wavelet transform. LH and HL focus on extracting edge features in the horizontal and vertical directions, respectively, while HH focuses on capturing diagonal texture information. Each high-frequency sub-band is then subjected to a 1×1 convolution to compress the channel dimension and reduce noise. Finally, the aggregated high-frequency features F are obtained by concatenating along the channel dimension. HF :

[0065] F HF =Concat(Conv1×1(LH), Conv1×1(HL), Conv1×1(HH)) (1)

[0066] Because high-frequency components are sparse and easily interfered with by noise, for F HF First, 2×2 max pooling is performed to preserve significant responses, followed by 3×3 depthwise separable convolution and 2×2 transposed convolution to obtain the intermediate high-frequency representation F. HF’ Then for F HF’ The spatially gated graph M is obtained by applying the Sigmoid activation function:

[0067] F HF’ =Upsampling(DWConv(MaxPool(F HF (2)

[0068] M = Sigmoid(F) HF’ (3)

[0069] The low-frequency representation is enhanced by element-wise weighting of the gated graph for the low-frequency subbands and summing the residuals:

[0070] L out =M⊙LL+LL (4)

[0071] The same high and low frequency processing modules are applied to the previous and next frames respectively to obtain three frames of enhanced low frequency features. Then, the features are upsampled to the original frame resolution through transposed convolution and fed into the ConvNeXt-T encoder with shared weights to generate a four-stage multi-scale frequency domain feature map.

[0072] Step 22 involves injecting temporal context into the frequency domain features of the fourth stage through a bidirectional cross-attention module, thereby obtaining the time-related frequency domain features C′4 at scale one. Specifically, this includes the following steps:

[0073] like Figure 4 As shown, the encoded frequency domain feature of the fourth stage of the current frame is C4. The key and value are first obtained through global average pooling and linear projection:

[0074] K = Re(W) k ·GAP(C4 (5)

[0075] V = Re(W) v ·GAP(C4 (6)

[0076] K and V are reshaped into 1×C to match subsequent dimensions. The fourth-stage features P4 and N4 of the previous and next frames are also subjected to global average pooling and linear projection to obtain the query vector Q1 = W. q1 GAP(P4) and Q2 = W q2 • GAP(N4), in channel attention, the channel similarity matrix A1 can be expressed as:

[0077]

[0078] The sum of each row is 1. And element (A1) jk This represents the similarity between channel j of the previous frame and channel k of the current frame. The vector is obtained by summing along the columns of A1. Then, perform element-wise multiplication with the value vector V and pass it through Sigmoid activation to obtain the forward channel weights:

[0079]

[0080] Similarly, for constructing the backward channel attention, the backward channel weight α2 guided by the next frame can be obtained only through C4 and the frequency domain feature N4 of the next frame output from the fourth stage. Finally, the forward and backward weights are adaptively fused:

[0081] α=λ1α1+λ2α2 (9)

[0082] stλ1+λ2=(10)

[0083] Where λ1 and λ2 are learnable parameters, the shape of the fusion weight α is reshaped as follows: The reconstructed temporally correlated frequency domain feature C′4 is obtained by multiplying it channel-by-channel with the current frame feature C4. This invention establishes channel-level dependencies between consecutive frames using a bidirectional channel cross-attention module, adaptively enhancing the features of important channels in the current frame while suppressing redundant channels, thereby achieving fine reconstruction of temporally correlated features. This module not only fully utilizes the temporal information of consecutive frames to improve feature consistency but also maintains low computational complexity through channel-level attention calculation.

[0084] Step 23: A lightweight cross-domain modulation module is introduced to achieve cross-domain intermodulation at scale one, and the modulated features are upsampled as input at scale two, such as... Figure 2 As shown, the cross-domain modulation module at scale one uses spatial feature S4 as the source input F. src Frequency domain feature C′4 is used as target input F tar Specifically, the steps include the following:

[0085] First, regarding F src Max pooling and average pooling are performed along the channel dimension, then concatenated along the channels and a 1×1 convolution is performed to restore the channel dimension to 1, resulting in the intermediate representation.

[0086]

[0087] This module considers both significant and average responses, providing a more comprehensive description of spatial importance. Next, it... Applying a linear transformation and using the Sigmoid activation function, we obtain the source domain spatial attention map S. att :

[0088]

[0089] Where w and b are learnable parameters. The attention map is multiplied element by element by the target input F. tar Adding residual connections to preserve connections in order to obtain the modulated target features F tar ′:

[0090] F tar ′=S att ⊙F tar +F tar (13)

[0091] The lower-level cross-domain feature module swaps the source and target (with C′4 as the source and S4 as the target) to obtain the modulated spatial representation S3.

[0092] Step 24: At scale two, channel-level reconstruction is first performed using C3 and the ConvNeXt-T frequency domain features P3 and N3 from the third stage of the preceding and following frames to obtain the reconstructed frequency domain feature C′3. Subsequently, two sets of cross-domain feature modulation modules are applied at this scale to intermodulate C′3 with the modulated spatial feature S3, thereby achieving cross-domain information exchange at a larger scale. Through cascaded interaction at two scales, temporal information is deeply integrated into the frequency domain representation of the current frame, and the complementary information across semantic levels enhances the discriminative ability and robustness of features in each domain.

[0093] Step 25: The final spatial and frequency domain features are converged using a fusion network, and a crowd density map for the current frame is generated using a regression head. This specifically includes the following steps:

[0094] First, the spatial and frequency domain features obtained after the second-scale interaction are upsampled to 1 / 8 of the resolution relative to the input and concatenated along the channel dimension. Second, three convolutional layers are used to progressively reduce the number of channels and fuse the information, resulting in the following structure:

[0095] C(256,3,1)-GC(128,3,1)-GC(64,3,1)

[0096] Where C(i, j, k) represents a convolutional layer with i kernels, a kernel size of j×j, and an inflation rate of k, and G is the GeLU activation function. Finally, the fused features are fed into a regression head to generate the final density map. The regression head is configured as follows:

[0097] C(32,3,1)-GC(1,1,1)

[0098] Step 3: Train a cascaded cross-domain feature interaction network;

[0099] Specifically, the steps include the following:

[0100] For the ConvNeXt-T model, weights pre-trained on ImageNet-22K are used for initialization. The optimizer for training the cascaded cross-domain feature interaction network is Adam, with an initial learning rate of 1e. - 5. The weights are decayed to 1e-4, the maximum number of iterations is 500, and the model weights that perform best on the validation set are saved.

[0101] Step four: Test the cascaded cross-domain feature interaction network. This includes the following steps:

[0102] All video frames in the test set are divided into groups of three consecutive frames and sequentially input into the optimal cascaded cross-domain feature interaction network model to obtain a density map P of consecutive T frames output by the density regression head. d Subsequently, regarding P d The model estimates the number of people in each test video frame by summing the results frame by frame.

[0103] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A video crowd counting method based on a cascaded cross-domain feature interaction network, characterized in that, Includes the following steps: S1 divides all frames of the video into training, validation, and test sets according to timeline and a certain ratio, ensuring that the video frames in each set are continuous. Data augmentation is achieved by applying uniform random cropping and horizontal flipping operations to each current frame and the two frames before and after it in the training set, thus obtaining an expanded training sample set. Missing frames at the beginning and end are directly filled in by copying the first and last frames. S2, a dual-branch feature coding network is constructed, comprising a frequency domain branch and a spatial domain branch. The frequency domain branch extracts the frequency domain features of each frame based on the image information of the current frame and its preceding and following frames through a high- and low-frequency signal aggregation module, and uses a feature encoder to obtain the multi-stage frequency domain feature representation of each frame. The spatial domain branch takes the image of the current frame as input and extracts spatial static features through a feature encoder. The frequency domain branch and the spatial domain branch are in parallel structure, providing feature input for subsequent cross-domain feature fusion. S3. A cascaded cross-domain feature interaction decoding network is constructed. This network first uses the current frame as the center, employing frequency domain features from previous and subsequent frames as auxiliary inputs, and extracts the temporally correlated frequency domain features of the current frame through bidirectional channel cross-attention. Subsequently, the extracted temporally correlated frequency domain features and the spatial domain static features of the current frame are input into the cross-domain feature intermodulation module to achieve mutual fusion and enhancement of the two domain features. The above temporally correlated frequency domain feature extraction and cross-domain feature intermodulation processes are repeatedly executed in a cascaded manner at different stages of the encoded features to gradually strengthen the semantic representation of the temporally correlated frequency domain and spatial domain of the current frame. Finally, the enhanced dual-branch features are concatenated, and channel compression and feature integration are performed through a multi-layer convolutional fusion network to output the crowd density map of the current frame. The estimated number of people in the current frame is obtained through global summation. S4. After multiple rounds of training on a cascaded cross-domain feature interaction network that simultaneously includes encoding and decoding, the network is tested on a validation set. The network weights corresponding to the best results on the validation set are saved to obtain the optimal network model. Each current frame and the two frames before and after it are used as samples in the test set and input into the optimal network model to estimate the number of people in each current frame.

2. The video crowd counting method based on a cascaded cross-domain feature interaction network according to claim 1, characterized in that, In step S2, the high and low frequency signal aggregation module performs frequency domain transformation on the input image to obtain high and low frequency sub-bands; it generates a spatial gating map using the local saliency of high frequency features, and uses the gating map to perform weighted enhancement and reconstruction of low frequency features to obtain a frequency domain feature representation of fused high and low frequency signals.

3. The video crowd counting method based on a cascaded cross-domain feature interaction network according to claim 1, characterized in that, In step S3, the bidirectional channel cross-attention module takes the current frame as the center and the adjacent frames before and after as auxiliary to construct the forward channel attention and the backward channel attention respectively. Then, it uses an adaptive weighting method to fuse the attention of the two channels to construct the temporal correlation frequency domain features of the current frame. The construction of the forward channel attention includes: using the frequency domain features of the current frame and the previous frame, respectively, to obtain the query vector of the previous frame and the key and value vectors of the current frame through global average pooling and linear transformation; calculating the attention matrix through the query vector of the previous frame and the key vector of the current frame and summing it in the first dimension; and then multiplying it by the value vector of the current frame and performing nonlinear activation to obtain the forward channel weight vector. The construction of the backward channel attention is similar to that of the forward channel attention, and the backward channel weight vector is calculated using the frequency domain features of the current frame and the frequency domain features of the next frame. The adaptive weighting method refers to the weighting and fusing of the forward and backward channel weight vectors according to learnable parameters λ1 and λ2 to obtain a vector α reflecting the weight of each channel, satisfying λ1+λ2=1; the vector α is multiplied with the frequency domain features of the current frame channel by channel to obtain the frequency domain features after temporal channel reconstruction.

4. The video crowd counting method based on a cascaded cross-domain feature interaction network according to claim 1, characterized in that, In step S3, the cross-domain feature mutual modulation module is used to realize the mutual modulation and enhancement of spatial domain features and frequency domain features based on time-series correlation reconstruction; This module uses one feature as the source domain and the other as the target domain. It performs max pooling and average pooling operations in parallel on the source domain features and fuses the pooling results to extract salient features from the source domain. Subsequently, it applies a linear transformation and nonlinear activation to these salient features, generating a spatial weight map of the source domain to modulate the target domain features, thus achieving cross-domain feature enhancement and information fusion. The source and target domains are interchangeable, resulting in frequency domain features enhanced by spatial domain modulation and spatial domain features enhanced by frequency domain modulation, respectively.

Images