Low-light environment tumble detection method based on adaptive image enhancement
Adaptive image enhancement using the DDPG algorithm solves the image quality and detection accuracy issues of fall detection in low-light environments, achieving efficient and real-time fall detection.
Patent Information
- Application Number
- CN202510978993.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-16
- Publication Date
- 2025-10-03
AI Technical Summary
Existing technologies for fall detection in low-light environments have problems such as limited image quality improvement, high computational complexity, high false alarm and missed alarm rates, and difficulty in achieving real-time detection.
The Deep Deterministic Policy Gradient (DDPG) algorithm is used for adaptive image enhancement. By dynamically estimating lighting conditions and optimizing image parameters, fall detection is performed in combination with the YOLOv5 model.
The accuracy and stability of fall detection in low-light environments are significantly improved, while maintaining real-time performance and adapting to different lighting and background disturbances.
Smart Images
Figure CN120748044A_ABST
Abstract
Description
[0001] FIELD OF THE INVENTION The present invention relates to the fields of intelligent monitoring, image processing, and computer vision technologies, and more particularly to an innovative method for improving fall detection accuracy in low-light environments through adaptive image enhancement technology. Background Art
[0002] Existing methods for fall detection in low-light environments mainly include traditional motion detection methods such as background subtraction, frame differencing, and optical flow, as well as histogram equalization (P. H¨artinger and C. Steger, “Adaptive histogram equalization in constant time,” Journal of Real-Time Image Processing, vol. 21, no. 3, p. 93, 2024.), Retinex theory and its variants (F. Tian, M.Wang, and X. Liu, “Low-light mine image enhancement algorithm based on im-\nprovedretinex,” Applied Sciences, vol. 14, no. 5, p. 2213, 2024.), and dark channel priors (G. Yang, H. Yang, S. Yu, J. Wang, and Z. Nie, “A multi-scale dehazing network with dark\nchannel priors,” Sensors, vol. 23, no. 13, p. 5980, 2023.) and other image enhancement technologies. However, these methods have many limitations when facing low-light environments, such as high computational complexity, large training data requirements, high system cost, and limited improvement in image quality.
[0003] The main shortcomings of existing technologies are: First, there is a lack of adaptive image enhancement technology for low-light environments, which cannot effectively improve image quality; second, fall detection algorithms are easily affected by interference in low-light conditions, resulting in high false alarm and missed alarm rates; third, traditional methods generally have high computational complexity, making real-time detection difficult. For example, background subtraction methods suffer from poor image quality and difficulty in background modeling in low-light environments, leading to increased error rates. Frame differencing methods struggle to accurately distinguish true motion from random fluctuations in low light and are severely affected by noise interference. Optical flow methods blur image details in low light, resulting in unstable optical flow calculations and affecting fall detection effectiveness. Traditional image enhancement techniques are mostly based on fixed transformation rules and lack accurate prior knowledge. They cannot dynamically respond to specific lighting conditions and are unable to flexibly adapt to diverse application scenarios. Therefore, the performance of existing technologies in fall detection in low-light environments is limited, failing to meet the high accuracy and real-time requirements of practical applications.
[0004] The technical problem that the present invention aims to solve is: to provide a fall detection method in low-light environments based on adaptive image enhancement, which improves the visual effect of images in low-light environments by dynamically adjusting image enhancement parameters, thereby improving the accuracy and stability of fall detection while maintaining real-time performance. Summary of the Invention
[0005] This paper proposes a low-light fall detection method based on adaptive image enhancement. This method aims to address the significant decline in accuracy of RGB-frame-based fall detection methods in low-light conditions. By combining the Deep Deterministic Policy Gradient (DDPG) algorithm with advanced image enhancement techniques, this method dynamically estimates and optimizes lighting conditions, significantly improving fall detection performance in low-light environments.
[0006] The technical solution of the present invention:
[0007] A fall detection method in low-light environments based on adaptive image enhancement, the steps are as follows:
[0008] (1) Feature extraction and lighting condition estimation:
[0009] First, key attributes are extracted from the original low-light image, including overall brightness, local contrast, and color balance. The DDPG model is used to learn the state space and action space. The state space contains parameters that characterize the lighting characteristics, while the action space includes possible enhancement measures. The actor network outputs the optimal action, and the critic network evaluates the impact of the action on future rewards, thereby accurately estimating the lighting conditions.
[0010] Furthermore, the overall brightness calculation formula is as follows:
[0011]
[0012] Among them, M and N represent the height and width of the image respectively, and Represents a pixel Grayscale value or color intensity.
[0013] Furthermore, the local contrast is defined as:
[0014]
[0015] in, In pixels The average brightness within a small area centered on the
[0016] Furthermore, the calculation formula of the color balance vector is:
[0017]
[0018] in, is the average value of the red channel pixel value in the image, is the average value of the green channel, is the average value of the blue channel, is the overall average value of the three channels;
[0019] Furthermore, the state space and action space The definition is as follows:
[0020]
[0021]
[0022] in, Control the brightness adjustment range, Control the local contrast enhancement coefficient, Controls color balance adjustment parameters.
[0023] Furthermore, the Actor network parameter update and Critic network loss in the DDPG model are defined as:
[0024]
[0025]
[0026]
[0027]
[0028] The parameters in the above formulas are explained as follows:
[0029] is the parameter set of the Actor network (i.e., the weight of the policy network);
[0030] is the parameter set of the Critic network;
[0031] The strategy objective function, i.e. the expected cumulative return;
[0032] Represents the action output by the Actor network
[0033] Outputs the gradient of the Actor network to its parameters for updating the policy network
[0034] Represents the output of the current Critic network, indicating that in state Next action the estimated total return that can be obtained;
[0035] is the expectation operator, which represents the expectation of the next state s' that follows the distribution P. P is p(s'|s,a), which represents the probability distribution of transitioning to the next state s' after executing action a in the current state s.
[0036] is the immediate reward obtained after executing action a in state s;
[0037] is the action output by the Actor network in state s ;
[0038] is the partial derivative of the Critic network with respect to the input action, indicating the impact of policy changes on returns;
[0039] Calculate the target value of the Critic network; is the actual reward obtained in step i, which is determined by the detection performance after image enhancement;
[0040] Discount factor, used to balance the contribution of current rewards and future rewards, here the value is 0.99;
[0041] The target critic network is used for stable training;
[0042] Generate the next state action for the target Actor network, keeping the update smooth;
[0043] Represents the Critic network loss, where N represents the batch size, that is, the number of samples in the batch.
[0044] (2) Adaptive image enhancement:
[0045] Based on the motion vectors output by the DDPG model, the expected lighting parameters are inferred. These parameters are used to guide the image enhancement process, including adjusting brightness, enhancing contrast, and correcting color deviations.
[0046] Furthermore, based on the Actor output action vector , the image enhancement module applies the following transformation:
[0047]
[0048] The formula controls brightness enhancement, contrast enhancement and color gain respectively. The enhanced image It is fed into the YOLOv5 model for fall detection.
[0049] The image enhancement reward function is defined as follows:
[0050]
[0051] in For accuracy, is the recall rate, is the mean average precision, is the weighting coefficient.
[0052] (3) Fall detection:
[0053] A pre-trained YOLOv5 object detection model is used for human posture recognition and fall status detection. The YOLOv5 model first quickly scans the enhanced image, locates human body parts, and annotates the position and size of the bounding box. It then analyzes the body's posture characteristics to trigger a fall alarm.
[0054] The beneficial effects of the present invention are: first, a specific implementation method of the adaptive image enhancement technology, especially the illumination condition estimation and image enhancement parameter optimization based on the DDPG model; second, the reward function construction combined with the detection accuracy feedback; third, the integrated enhancement-detection collaborative structure with the YOLOv5 detection module. Compared with the existing technology, the technical advantages of the present invention are mainly reflected in the following aspects: first, it has an end-to-end trainable structure, and the enhancement and detection are collaboratively optimized; second, the enhancement parameters are dynamically adjusted, which is better than the fixed enhancement method; third, it has real-time response capabilities while maintaining high detection accuracy; fourth, it is more robust to different illumination and background disturbances. The present invention is not only technically innovative, but also has achieved remarkable results in practical applications. With the growing demand for intelligent monitoring and health monitoring of the elderly, this method is expected to become one of the standard technical solutions in related fields. The low-light environment fall detection method based on adaptive image enhancement proposed in the present invention not only solves the problems existing in the existing technology, but also shows a broad application prospect and potential standard setting value. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 It is a schematic diagram of the technical process of the present invention;
[0056] Figure 2 Schematic diagram of feature extraction;
[0057] Figure 3 Schematic diagram of the DDPG training process. DETAILED DESCRIPTION
[0058] This proposal addresses the challenges of fall detection in low-light environments by proposing an adaptive image enhancement solution combined with the Deep Deterministic Policy Gradient (DDPG) algorithm. This solution improves image quality, effectively enhancing the accuracy and reliability of fall detection. The following is a detailed description of the technical solution:
[0059] 1. Dataset Construction and Distribution
[0060] To ensure comprehensiveness and representativeness of the experiment, a dataset of RGB images of both fall and non-fall events in low-light environments was constructed. This dataset covers a variety of scenarios (e.g., indoors and outdoors), varying lighting conditions (e.g., nighttime and daytime), and a variety of fall actions to truly reflect the diversity of the world. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio to ensure sufficient sample size for the model during training, while retaining independent data for validation and final evaluation of model performance.
[0061] 2. Feature Extraction
[0062] Feature extraction is the first step in the entire pipeline, which extracts key attributes from the original low-light image for subsequent processing.
[0063] The feature extraction in this embodiment is as follows Figure 2 As shown in the figure, a CNN network is used. The network design uses a gradually increasing number of channels and a gradually decreasing feature map size to extract hierarchical features. The specific architecture is as follows:
[0064] First, the original input image, with a size of 256×256×3, passes through the first convolutional layer. This convolutional layer uses a 3×3 convolution kernel with a stride of 2, outputting a 32-channel feature map with a size of 128×128×32. A ReLU activation function is then applied. This feature map is then further compressed to 64×64×32 through the first max pooling layer (2×2 with a stride of 2), reducing the spatial dimensionality of the feature map while retaining the most important feature information.
[0065] The data then enters the second convolutional layer, which also uses a 3×3 convolution kernel with a stride of 1, outputting a 64-channel feature map of size 64×64×64. A ReLU activation function is applied. The feature map is then compressed to 32×32×64 by passing through a second max pooling layer (2×2 with a stride of 2).
[0066] Next comes the third convolutional layer, which uses a 3×3 convolution kernel with a stride of 1. The number of output channels increases to 128, while the feature map size remains at 32×32×128. ReLU activation function is also applied. This layer extracts higher-level semantic features.
[0067] After the convolution feature extraction is completed, global average pooling is used to compress the 32×32×128 feature map into a 1×1×128 feature vector. This method can not only retain important global information, but also significantly reduce the number of parameters and avoid overfitting.
[0068] Finally, feature mapping is performed through two fully connected layers: the first fully connected layer contains 512 neurons, uses the ReLU activation function and applies the Dropout technique to prevent overfitting; the second fully connected layer outputs a 3D feature vector corresponding to the state space s_t = [L_avg, Con, CB], which represents the three key lighting features of overall brightness, local contrast and color balance respectively.
[0069] This design achieves a progressive extraction from low-level edge texture features to high-level semantic features by gradually increasing the number of channels (32 → 64 → 128) and gradually reducing the spatial size. Ultimately, it compresses complex image information into a three-dimensional illumination feature vector, providing an effective state representation for subsequent DDPG reinforcement learning. The following key properties are obtained:
[0070] (1) Overall brightness :
[0071] The overall brightness reflects the average brightness of the entire image, and the calculation formula is as follows:
[0072]
[0073] Among them, M and N represent the height and width of the image respectively, and Represents a pixel Grayscale value or color intensity.
[0074] (2) Local contrast ( , ):
[0075] Local contrast measures the changes in brightness and darkness within a local area of an image, which helps capture edges and other details. It is defined as:
[0076]
[0077] here, In pixels The average brightness within a small area centered on the
[0078] (3) Color balance ( , FF1A
[0079] Color balancing is used to evaluate the distribution of different colors in an image, ensuring that the relative intensities of each channel remain consistent. This is achieved by calculating a color histogram, which adjusts the color effects of the image to make it more similar to natural visual perception.
[0080]
[0081] in, is the average value of the red channel pixel value in the image, is the average value of the green channel, is the average value of the blue channel, is the overall average value of the three channels;
[0082] The above features constitute the state vector input to the DDPG model, that is,
[0083]
[0084] 3. Definition of state space and action space
[0085] In the present invention, the state space Characterizes the illumination and perceptual characteristics of the current input image, action space Characterizes the enhancement operation to be applied to the image and is defined as follows:
[0086]
[0087]
[0088] in, Control the brightness adjustment range, Control the local contrast enhancement coefficient, Controls color balance adjustment parameters.
[0089] 4. Reward Function Design
[0090] In order to make the image enhancement results helpful for subsequent fall detection, we tightly couple the reward function with the detector performance, which is defined as follows:
[0091]
[0092] in For accuracy, is the recall rate, is the mean average precision.
[0093] 5.DDPG training process
[0094] The training adopts the typical DDPG framework, and the Actor network parameter update and Critic network loss are defined as:
[0095]
[0096]
[0097]
[0098] in Calculate the target value of the Critic network.
[0099] 6. Image enhancement and detection joint process:
[0100] Output motion vector based on Actor , the image enhancement module applies the following transformation:
[0101]
[0102] This formula controls brightness enhancement, contrast enhancement, and color gain, respectively. The enhanced image is then fed into the YOLOv5 model for human detection. The final fall determination is based on a comprehensive analysis of the detection frame position, posture angle, and temporal continuity rules, achieving collaborative optimization of enhancement and detection, improving detection accuracy and robustness.
[0103] 7. Experimental evaluation and result analysis
[0104]
[0105] To objectively evaluate the effectiveness of the proposed method, we used three key metrics: precision, recall, and mean average precision (mAP) for quantitative analysis. Experimental results show that with increasing training epochs, all metrics show an upward trend for both original low-light images and enhanced images. In particular, the model reached a good convergence state after the 250th epoch.
[0106] Compared to the original low-light images, the images enhanced using the DDPG algorithm show significant improvements across all evaluation metrics. For example, after 300 epochs, the detection accuracy of the enhanced images reached 90.2%, 3.7 percentage points higher than the original images; and the mean average precision (MAP) increased from 85.3% to 87.8%, a 2.5 percentage point increase. This fully demonstrates the effectiveness of this solution in improving fall detection performance in low-light environments.
[0107] 8. Innovations and advantages of technical solutions
[0108] The innovation of this technical solution lies primarily in the personalized enhancement of images achieved through the DDPG algorithm, significantly improving image quality. While maintaining model complexity and computational efficiency, it effectively improves the accuracy and reliability of fall detection. This solution is not only applicable to the field of fall detection, but can also be extended to other application scenarios requiring image enhancement. The adaptive image enhancement technology solution combined with the DDPG algorithm proposed in this proposal demonstrates excellent performance in solving the problem of fall detection in low-light conditions. This solution is not only theoretically innovative, but also demonstrates good practicality and promotion potential in practical applications.
[0109] As can be seen from the table, with the increase in training rounds, all indicators show an upward trend, both for the original low-light images and the enhanced images. In particular, after the 250th epoch, the precision P, recall R, and mAP are basically stable, indicating that the model has reached a good state of convergence.
[0110] It is particularly noteworthy that compared to the original low-light images, the images enhanced by the DDPG algorithm have been significantly improved in all evaluation indicators. For example, at 300 epochs, the detection accuracy of the enhanced image reached 90.2%, which is 3.7 percentage points higher than the original image; and the mAP also increased from 85.3% to 87.8%, an increase of 2.5 percentage points. This fully demonstrates the effectiveness of the method based on adaptive image enhancement proposed in this paper in improving the performance of fall detection in low-light environments. Through detailed experimental settings and complex data analysis, it can be concluded that the adaptive image enhancement scheme combined with the DDPG algorithm proposed in this study can not only provide theoretical support in theory, but also show excellent performance in practical applications, providing new ideas and technical means for solving the problem of fall detection in low-light conditions.
[0111] Despite employing a complex deep learning framework, the system's real-time responsiveness is ensured by optimizing the network architecture, employing efficient GPU-accelerated computing, and implementing a reasonable batch processing strategy. Furthermore, the method demonstrates strong robustness, maintaining stable detection performance even in the face of extreme lighting changes or complex background interference.
Claims
1. A fall detection method in low-light environment based on adaptive image enhancement, characterized in that: Here are the steps: (1) Feature extraction and lighting condition estimation: First, key attributes are extracted from the original low-light image, including overall brightness, local contrast, and color balance. The DDPG model is used to learn the state space and action space. The state space contains parameters that characterize lighting characteristics, while the action space covers possible enhancement measures. The actor network outputs the optimal action, and the critic network evaluates the impact of the action on future rewards, thereby accurately estimating lighting conditions. (2) Adaptive image enhancement: Based on the motion vectors output by the DDPG model, we infer the expected lighting parameters. These parameters are used to guide the image enhancement process, including adjusting brightness, enhancing contrast, and correcting color deviations. (3) Fall detection: A pre-trained YOLOv5 object detection model is used for human posture recognition and fall status judgment. The YOLOv5 model first quickly scans the enhanced image, locates the human body parts, and marks the position and size of the bounding box. It then analyzes the human posture characteristics to trigger a fall alarm.
2. The method for fall detection in low-light environment based on adaptive image enhancement according to claim 1, characterized in that: In step (1), the overall brightness calculation formula is as follows: ， Among them, M and N represent the height and width of the image respectively, and Represents a pixel Grayscale value or color intensity.
3. The method for fall detection in low-light environment based on adaptive image enhancement according to claim 1, characterized in that: In step (1), the local contrast is defined as: ， in, In pixels The average brightness within a small area centered on the 4. The method for fall detection in low-light environment based on adaptive image enhancement according to claim 1, characterized in that: In step (1), the calculation formula of the color balance vector is: ， in, is the average value of the red channel pixel value in the image, is the average value of the green channel, is the average value of the blue channel, is the overall average of the three channels.
5. The method for fall detection in low-light environment based on adaptive image enhancement according to claim 1, characterized in that: In the step (1), the state space and action space The definition is as follows: ， ， in, Control the brightness adjustment range, Control the local contrast enhancement coefficient, Controls color balance adjustment parameters.
6. The method for fall detection in low-light environment based on adaptive image enhancement according to claim 1, characterized in that: In step (1), the Actor network parameter update and Critic network loss in the DDPG model are defined as: ， ， ， ， The parameters in the above formulas are explained as follows: is the parameter set of the Actor network, that is, the weight of the policy network; is the parameter set of the Critic network; The strategy objective function, i.e. the expected cumulative return; Represents the action output by the Actor network Outputs the gradient of the Actor network to its parameters for updating the policy network Represents the output of the current Critic network, indicating that in state Next action the estimated total return that can be obtained; is the expectation operator, which means the expectation of the next state s' that obeys the distribution P; where P is p(s'|s,a), which represents the probability distribution of transitioning to the next state s' after executing action a in the current state s; is the immediate reward obtained after executing action a in state s; is the action output by the Actor network in state s ; is the partial derivative of the Critic network with respect to the input action, indicating the impact of policy changes on returns; Calculate the target value of the Critic network; is the actual reward obtained in step i, which is determined by the detection performance after image enhancement; Discount factor, used to balance the contribution of current rewards and future rewards, here the value is 0.99; The target critic network is used for stable training; Generate the next state action for the target Actor network, keeping the update smooth; Represents the Critic network loss, where N represents the batch size, that is, the number of samples in the batch.
7. The method for fall detection in low-light environment based on adaptive image enhancement according to claim 1, characterized in that: In the step (2), based on the Actor output action vector , the image enhancement module applies the following transformation: ， The formula controls brightness enhancement, contrast enhancement and color gain respectively. The enhanced image It is fed into the YOLOv5 model for fall detection; The image enhancement reward function is defined as follows: ， in For accuracy, is the recall rate, is the mean average precision, is the weighting coefficient.