Vehicle-side collaborative panoramic monitoring method based on reinforcement learning
By segmenting the panoramic image into sub-images and processing it in parallel, combining intelligent decision-making of edge computing and reinforcement learning, the challenges of computing burden and real-time requirements in panoramic image object detection are solved, and efficient and accurate panoramic object detection is achieved.
Patent Information
- Application Number
- CN202510171249.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-17
- Publication Date
- 2025-06-06
AI Technical Summary
The existing deep learning-based panoramic image object detection methods have challenges in data processing and real-time requirements, including problems such as computing resource allocation, data transmission bandwidth and real-time requirements.
A vehicle-side collaborative panoramic monitoring method based on reinforcement learning is proposed. By segmenting the panoramic image into multiple sub-images and processing it in parallel, combining edge computing and distributed parallel computing, intelligent offload decision and binary graph matching algorithm are adopted to improve detection efficiency and accuracy.
It effectively reduces the computing burden and delay, improves the efficiency and accuracy of object detection, meets the real-time requirements, and optimizes the utilization of system resources.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] The present application belongs to the field of intelligent information processing and relates to a vehicle-side collaborative panoramic monitoring method based on reinforcement learning. Background Art
[0002] Omnidirectional visual perception technology has been widely used in the Internet of Things environment, especially in the fields of virtual reality, autonomous driving, and intelligent robots. It has become a key perception technology because it can provide panoramic views and high-resolution images. As the core component of this technology, panoramic cameras can capture images with a 360-degree perspective and provide extremely rich scene information. However, the special nature of panoramic images also brings many challenges, mainly in terms of image distortion, high computational complexity, and high real-time requirements. Especially in visual tasks such as target detection, how to overcome these problems while ensuring detection accuracy is still a current research difficulty.
[0003] In recent years, some researchers have tried to design new network architectures or modules based on the characteristics of panoramic images to better cope with the challenges of panoramic object detection. However, existing deep learning-based panoramic image object detection methods still face many challenges, especially in terms of data processing and real-time requirements. First, due to the wide field of view of panoramic images, foreground objects are unevenly distributed in the image, resulting in dense objects in some areas and sparse objects in other areas, which brings additional difficulties to object detection. Secondly, panoramic images usually have a high resolution, which causes high latency in image transmission and processing, making it difficult to meet the needs of real-time detection. In addition, due to the scarcity of panoramic image datasets, deep learning-based object detection algorithms lack sufficient training samples, which limits the generalization ability and robustness of the algorithm.
[0004] With the advent of 5G communication technology, edge computing and distributed parallel computing have gradually become key technologies for solving computationally intensive tasks and large-scale model training. Edge computing can effectively reduce transmission delays and computing burdens by migrating data processing tasks from the cloud to edge devices close to the data source. However, existing edge computing methods still face a series of challenges in panoramic image object detection tasks, such as computing resource allocation, data transmission bandwidth, and real-time requirements. Therefore, how to combine the advantages of edge computing and distributed parallel computing to design an efficient panoramic image object detection method is still a key issue that needs to be solved urgently. Summary of the invention
[0005] In order to overcome the above defects, this application proposes a vehicle-side collaborative panoramic monitoring method based on reinforcement learning. The specific steps of this application are as follows:
[0006] S1, divides the panoramic image into multiple sub-images and transmits them to the edge device for parallel processing through the optimal overlapping domain strategy to improve detection efficiency;
[0007] S2, the edge device independently processes the assigned sub-image and performs preliminary object detection, reducing computational burden and latency;
[0008] S3, based on the system status information, intelligently decides whether to upload the sub-image to the cloud for further processing;
[0009] S4, using a bipartite graph matching algorithm to eliminate redundant detection frames in the overlapping areas of adjacent sub-images to ensure the accuracy and consistency of the final detection results;
[0010] S5: output accurate panoramic target detection results.
[0011] The technical features and improvements of this application are:
[0012] For step S1, the present application divides the panoramic image into multiple sub-images according to a preset size. The design of each sub-image includes an appropriate overlapping area to ensure that the overlapping part between adjacent sub-images is large enough to avoid the target being cut at the splitting boundary and prevent missed detection. The split sub-images are sent to the edge computing device for processing by parallel transmission, which reduces the computing burden of a single device and improves the overall processing efficiency. The design of the overlapping area adopts the optimal overlapping domain strategy to balance the computing burden and the accuracy of target detection. By calculating the width s of each sub-image and the width r of the overlapping area, it is ensured that the target can be fully identified in the adjacent sub-images, and the target is avoided from being improperly cut during the splitting process, avoiding the occurrence of false detection or missed detection. Assuming that the width of the panoramic image is W, the width of the split sub-image is s, and the width of the overlapping area between the sub-images is r, then:
[0013] Num×s-(Num-1)×r=W(1)
[0014] Wherein, Num represents the number of sub-images, s represents the sub-image width, r represents the width of the overlapping area between sub-images, and W represents the total width of the panoramic image, ensuring that the width of the split sub-images and the size of the overlapping area meet the total width requirement of the panoramic image.
[0015] For step S2, after the panoramic image is split, the sub-images are transmitted to multiple edge computing devices. Each edge device independently processes its assigned sub-image and performs preliminary target detection. A deep convolutional neural network model is used to classify and locate potential targets in the sub-images. In view of the limited hardware resources of the edge devices, a lightweight model is adopted to accelerate calculations and reduce latency. Through parallel computing, the system can process multiple sub-images on multiple edge devices at the same time, thereby significantly reducing the overall computing time. Each edge device performs preliminary target detection on the split sub-images and generates corresponding detection boxes and confidence levels. Due to the small size of each sub-image, the edge device can quickly complete the detection task, ensuring that the system can respond in real time and meet the requirements of real-time processing.
[0016] For step S3, this application adopts an intelligent offloading decision method based on deep reinforcement learning, aiming to intelligently select which sub-images need to be uploaded to the cloud for further processing; specifically, the system dynamically decides whether to upload a sub-image to the cloud based on factors such as the vehicle's motion state, network bandwidth, device load, and task complexity; the core of the intelligent offloading decision is to make decisions based on the current system state, which includes the complexity of the image, the computing load of the edge device, and the current network status. The state space S t It can be expressed as:
[0017] S t =[c,n,v,e](2)
[0018] Where c represents the target detection confidence of the current sub-image, n is the color complexity of the image, v is the computational delay of the edge device, and e is the entropy value of the image, which represents the complexity of the image. Based on the current state S t , Intelligent Unloading Decision A t Outputted by the deep reinforcement learning model, its decision can be expressed as:
[0019]
[0020] Among them, A t =0 means not to upload to the cloud, A t =1 means uploading to the cloud; Q value function Q(S t ,A) is the long-term reward estimate calculated by the reinforcement learning model when executing action A under a given state; by continuously optimizing the offloading strategy, the system can dynamically adjust the offloading decision according to the real-time changing network conditions, edge computing resources and traffic environment, thereby achieving load balancing and bandwidth utilization optimization, ensuring efficient utilization of system resources and improving overall performance.
[0021] For step S4, the present application proposes a redundant frame elimination method based on a bipartite graph matching algorithm. After the target detection tasks are completed on the edge device and the cloud, due to the existence of overlapping areas of multiple sub-images, multiple detection frames may appear for the same target. Therefore, in order to merge the detection results and eliminate redundant frames to ensure that each target is detected only once, the present application adopts a bipartite graph matching algorithm. First, the matching cost between each pair of detection frames is calculated. The cost function comprehensively considers the position overlap of the target frame (through the intersection-over-union ratio, IOU) and the confidence of the detection frame. The matching cost M(a,b) is the matching cost of the target frames a and b, and the matching loss function is defined as:
[0022]
[0023] Where A and B are the detection box sets of the two sub-images, respectively, and M(a,b) is the matching cost between the target boxes a and b. The matching cost M(a,b) is mainly calculated based on the spatial overlap and confidence of the target boxes, and is usually defined by the intersection over union (IOU) and the confidence weight:
[0024] M(a,b)=λ conf (c a +c b )+λ iou IOU(a,b)(5)
[0025] Among them, λ iou and λ conf are weight coefficients, representing the contribution of IOR and confidence to the matching cost, c a and c b They represent the confidence of box a and box b respectively, and IOU represents intersection-over-union ratio, which measures the overlap of target boxes. Next, the optimal matching scheme is found in the bipartite graph through the Hungarian algorithm or the minimum cost matching method, and the redundant detection boxes in the overlapping area are eliminated to ensure that each target is detected only once. Through this matching algorithm, the system can efficiently select the optimal detection box to avoid repeated detection of the same target. After the redundant boxes are eliminated, the merged detection results ensure the accuracy and consistency of target detection. Finally, the detection boxes with low confidence or unreliable confidence are removed, and the boxes with the highest confidence are retained as the final detection results to ensure that the output detection boxes have high accuracy and credibility.
[0026] The vehicle-side collaborative panoramic monitoring method based on reinforcement learning in this application solves the problems of excessive computational burden, insufficient redundant data processing, and unintelligent offloading decision-making in panoramic image processing and target detection in the prior art, and has the following advantages:
[0027] (1) This application splits the panoramic image into multiple sub-images by adopting the optimal overlapping domain strategy and transmits them in parallel to the edge device for processing, effectively reducing the computational burden of each sub-image and improving the detection efficiency. At the same time, it ensures that the target is not cut during the splitting process to avoid missed detection problems.
[0028] (2) This application introduces an intelligent offloading decision-making mechanism based on deep reinforcement learning, which evaluates the system status in real time, including factors such as target detection confidence, edge device load, and network bandwidth, and intelligently chooses whether to upload the sub-image to the cloud for further processing; this method can reduce network bandwidth consumption and computing resource pressure while ensuring accuracy by optimizing the offloading path, thereby improving the overall efficiency of the system;
[0029] (3) This application uses a bipartite graph matching algorithm to eliminate redundant frames, effectively eliminating redundant detection frames in the overlapping areas between multiple sub-images, ensuring the accuracy and consistency of target detection results. When processing overlapping areas of images, the algorithm can avoid repeated detection to the greatest extent possible, ensuring that the final output panoramic target detection results have high accuracy;
[0030] (4) Through distributed parallel processing of edge computing devices, the system can simultaneously execute target detection tasks on multiple edge devices, significantly reducing processing time and latency, meeting real-time requirements, effectively reducing dependence on cloud computing resources, and improving the system's response speed and processing capabilities. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 This is a flow chart of the vehicle-side collaborative panoramic monitoring method in this application.
[0032] Figure 2 This is a diagram of the image segmentation method in this application.
[0033] Figure 3 This is a block diagram for eliminating redundant detection in bipartite graph matching in this application. DETAILED DESCRIPTION
[0034] The present application is further described in detail below with reference to the accompanying drawings and specific implementation methods:
[0035] A vehicle-side collaborative panoramic monitoring method based on reinforcement learning, such as Figure 1 As shown, it is a flow chart of the vehicle-side collaborative panoramic monitoring method based on reinforcement learning of the present application, and the method comprises:
[0036] S1. In order to effectively deal with the high resolution and complex computing problems of panoramic images, this application proposes an image segmentation method based on the optimal overlapping domain strategy. The panoramic image is divided into multiple sub-images to facilitate parallel processing on edge devices, thereby improving processing speed and reducing latency. The core of the optimal overlapping domain strategy is to ensure that there is enough overlapping area between adjacent sub-images to avoid the target being divided into two parts during the image segmentation process. Specifically, assuming that the width of the panoramic image is W, the width of the sub-image after segmentation is s, the width of the overlapping area between each sub-image is r, and the length of the maximum target bounding box is L max , then we have the following geometric relationship:
[0037] Num×s-(Num-1)×r=W(6)
[0038] Where Num is the number of sub-images after splitting. By optimizing the selection of r, each target can appear completely in a sub-image, avoiding missed detection or redundant detection due to target segmentation; the sub-images after image splitting are transmitted to each edge device in parallel, and each device handles its own task independently, thereby reducing data transmission time and improving real-time performance.
[0039] S2, the edge device performs preliminary target detection on the received sub-image through parallel computing. To reduce the computing burden, the edge device uses a lightweight target detection model. Each edge device independently completes the target detection task of the sub-image it is responsible for according to its computing resources and network bandwidth limitations. The basic form of the target detection model is:
[0040] Y i =f θ (X i )(7)
[0041] Among them, Y i represents the i-th sub-image X i The detection results on the target include the location, category and confidence level of the target, f θ It is a target detection model based on convolutional neural network, and θ is the parameter of the model. During the target detection process, the device not only outputs the bounding box of the target, but also outputs the category probability and confidence of each target. i For target i to belong to a certain category c k The probability of for:
[0042]
[0043] Among them, b i is the coordinate of the detection box, b gtis the true target box, IOU stands for intersection-over-union, which is used to measure the overlap between the detection box and the true box. The edge device filters out the qualified target box according to the confidence level and outputs it as the preliminary detection result.
[0044] S3, adopts an intelligent offloading decision method based on deep reinforcement learning, which aims to intelligently select which sub-images need to be uploaded to the cloud for further processing; specifically, the system dynamically decides whether to upload a sub-image to the cloud based on factors such as the vehicle's motion state, network bandwidth, device load, and task complexity; the core of intelligent offloading decision is to make decisions based on the current system state, which includes the complexity of the image, the computing load of the edge device, and the current network status. The state space S t It can be expressed as:
[0045] S t =[c,n,v,e](9)
[0046] Where c represents the target detection confidence of the current sub-image, n is the color complexity of the image, v is the computational delay of the edge device, and e is the entropy value of the image, which represents the complexity of the image. Based on the current state S t , Intelligent Unloading Decision A t Outputted by the deep reinforcement learning model, its decision can be expressed as:
[0047]
[0048] Among them, A t =0 means not to upload to the cloud, A t =1 means uploading to the cloud; Q value function Q(S t ,A) is the long-term reward estimate calculated by the reinforcement learning model when executing action A under a given state; by continuously optimizing the offloading strategy, the system can dynamically adjust the offloading decision according to the real-time changing network conditions, edge computing resources and traffic environment, thereby achieving load balancing and bandwidth utilization optimization, ensuring efficient utilization of system resources and improving overall performance.
[0049] S4, after the target detection results of multiple sub-images are summarized, there is a problem that multiple sub-images may perform redundant detection on the same target. In order to eliminate these redundant detection frames, the system uses an algorithm based on bipartite graph matching. The overlapping area between each two adjacent sub-images will generate redundant detection frames, so the redundancy of the overlapping part needs to be eliminated through an algorithm. Specifically, suppose sub-image X i and X j The target boxes detected in the overlapping area are a i and b i, we determine whether the two detection boxes belong to the same target by calculating their intersection over union (IOU). If the IOU value exceeds the set threshold, the two boxes are considered redundant and need to be merged. To this end, the problem of eliminating redundant detection boxes is transformed into a bipartite graph matching problem. A bipartite graph is constructed, where the detection box a of each sub-image i and b i As two sets of graphs, the matching cost between the two is determined by calculating the similarity or IOU between the boxes. The minimum cost matching is solved by the Hungarian algorithm or the maximum flow algorithm to eliminate redundant detection boxes. The matching cost formula is:
[0050]
[0051] Among them, C(a i ,b j ) is the matching cost, and IOU is usually used as the similarity metric. Through bipartite graph matching, we can effectively eliminate redundant boxes and ensure the accuracy of the final detection results.
[0052] S5, after completing the redundancy elimination of the target box, the system will output the final panoramic target detection result. The result not only contains the category information of each target, but also includes its location, confidence and other data. The final panoramic target detection result can be expressed by the following formula:
[0053]
[0054] Among them, Y i is the detection result of the ith sub-image, which contains all the target information in the sub-image. By removing redundant frames and merging the detection results of all sub-images, the system can obtain the final target detection result. Ensure completeness and consistency of target detection.
[0055] In summary, this application proposes a vehicle-side collaborative panoramic monitoring method based on reinforcement learning. This method improves the efficiency of target detection by dividing the panoramic image into multiple sub-images and transmitting them to the edge device through the optimal overlapping domain strategy; reduces the computing burden and delay through parallel computing and preliminary target detection on the edge device; dynamically selects which sub-images to upload to the cloud for further processing through an intelligent offloading decision model based on information value, thereby optimizing task offloading decisions; eliminates redundant detection frames through a bipartite graph matching algorithm to ensure the accuracy and consistency of the final result. This method effectively reduces the transmission of redundant data, optimizes the utilization of computing resources, and maintains efficient operation and long-term stability of the system in dynamic and complex traffic and network environments.
[0056] Although the content of the present application has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present application. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present application can be made. Therefore, the protection scope of the present application should be limited by the appended claims.
Claims
1. A vehicle-side collaborative panoramic monitoring method based on reinforcement learning, its characteristics and The specific steps are as follows: S1, divides the panoramic image into multiple sub-images and transmits them to the edge device for parallel processing through the optimal overlapping domain strategy to improve detection efficiency; S2, the edge device independently processes the assigned sub-image and performs preliminary object detection, reducing computational burden and latency; S3, based on the system status information, intelligently decides whether to upload the sub-image to the cloud for further processing; S4, using a bipartite graph matching algorithm to eliminate redundant detection frames in the overlapping areas of adjacent sub-images to ensure the accuracy and consistency of the final detection results; S5: output accurate panoramic target detection results.
2. According to the reinforcement learning-based vehicle-side collaborative panoramic monitoring method of claim 1, it is characterized in that: For step S1, the present invention divides the panoramic image into multiple sub-images according to a preset size. The design of each sub-image includes an appropriate overlapping area to ensure that the overlapping part between adjacent sub-images is large enough to avoid the target being cut at the splitting boundary and prevent missed detection. The split sub-images are sent to the edge computing device for processing by parallel transmission, which reduces the computing burden of a single device and improves the overall processing efficiency. The design of the overlapping area adopts the optimal overlapping domain strategy to balance the computing burden and the accuracy of target detection. By calculating the width s of each sub-image and the width r of the overlapping area, it is ensured that the target can be fully identified in the adjacent sub-images, and the target is prevented from being improperly cut during the splitting process, thereby avoiding the occurrence of false detection or missed detection. Assuming that the width of the panoramic image is W, the width of the split sub-image is s, and the width of the overlapping area between the sub-images is r, then: Num×s-(Num-1)×r=W (1) Wherein, Num represents the number of sub-images, s represents the sub-image width, r represents the width of the overlapping area between sub-images, and W represents the total width of the panoramic image, ensuring that the width of the split sub-images and the size of the overlapping area meet the total width requirement of the panoramic image.
3. The vehicle-side collaborative panoramic monitoring method based on reinforcement learning according to claim 1 is characterized in that: For step S2, after the panoramic image is split, the sub-images of the present invention are transmitted to multiple edge computing devices, and each edge device independently processes its assigned sub-image and performs preliminary target detection; a deep convolutional neural network model is used to classify and locate potential targets in the sub-images. In view of the limited hardware resources of the edge devices, a lightweight model is adopted to accelerate calculations and reduce delays; through parallel computing, the system can process multiple sub-images on multiple edge devices at the same time, thereby significantly reducing the overall calculation time; each edge device performs preliminary target detection on the split sub-images and generates corresponding detection boxes and confidence levels. Since the size of each sub-image is small, the edge device can quickly complete the detection task, ensuring that the system can respond in real time and meet the requirements of real-time processing.
4. The vehicle-side collaborative panoramic monitoring method based on reinforcement learning according to claim 1 is characterized in that: For step S3, the present invention adopts an intelligent offloading decision method based on deep reinforcement learning, aiming to intelligently select which sub-images need to be uploaded to the cloud for further processing; specifically, the system dynamically decides whether to upload a sub-image to the cloud based on factors such as the vehicle's motion state, network bandwidth, device load, and task complexity; the core of the intelligent offloading decision is to make decisions based on the current system state, which includes the complexity of the image, the computing load of the edge device, and the current network status. The state space S t It can be expressed as: S t =[c,n,v,e] (2) Where c represents the target detection confidence of the current sub-image, n is the color complexity of the image, v is the computational delay of the edge device, and e is the entropy value of the image, which represents the complexity of the image. Based on the current state S t , Intelligent Unloading Decision A t Outputted by the deep reinforcement learning model, its decision can be expressed as: Among them, A t =0 means not to upload to the cloud, A t =1 means uploading to the cloud; Q value function Q(S t ,A) is the long-term reward estimate when executing action A under a given state calculated by the reinforcement learning model; by continuously optimizing the offloading strategy, the system can dynamically adjust the offloading decision according to the real-time changing network conditions, edge computing resources and traffic environment, thereby achieving load balancing and bandwidth utilization optimization, ensuring efficient utilization of system resources and improving overall performance.
5. The vehicle-side collaborative panoramic monitoring method based on reinforcement learning according to claim 1 is characterized in that: For step S4, the present invention proposes a redundant frame elimination method based on a bipartite graph matching algorithm. After the target detection tasks are completed on the edge device and the cloud, due to the existence of overlapping areas of multiple sub-images, multiple detection frames may appear for the same target. Therefore, in order to merge the detection results and eliminate redundant frames to ensure that each target is detected only once, the present application adopts a bipartite graph matching algorithm. First, the matching cost between each pair of detection frames is calculated. The cost function comprehensively considers the position overlap of the target frame (through the intersection-and-union ratio, IOU) and the confidence of the detection frame. The matching cost M(a, b) is the matching cost of the target frames a and b, and the matching loss function is defined as: Where A and B are the detection box sets of the two sub-images, respectively, and M(a,b) is the matching cost between the target boxes a and b. The matching cost M(a,b) is mainly calculated based on the spatial overlap and confidence of the target boxes, and is usually defined by the intersection over union (IOU) and the confidence weight: M(a,b)=λ conf (c a +c b )+λ iou ·IOU(a,b) (5) Among them, λ iou and λ conf are weight coefficients, representing the contribution of IoU and confidence to the matching cost, c a and c b They represent the confidence of box a and box b respectively, and IOU represents intersection-over-union ratio, which measures the overlap of target boxes. Next, the optimal matching scheme is found in the bipartite graph through the Hungarian algorithm or the minimum cost matching method, and the redundant detection boxes in the overlapping area are eliminated to ensure that each target is detected only once. Through this matching algorithm, the system can efficiently select the optimal detection box to avoid repeated detection of the same target. After the redundant boxes are eliminated, the merged detection results ensure the accuracy and consistency of target detection. Finally, the detection boxes with low confidence or unreliable confidence are removed, and the boxes with the highest confidence are retained as the final detection results to ensure that the output detection boxes have high accuracy and credibility.