Traffic light control method with ambient edge intelligence based on edge computing

Abstract

The present invention provides a traffic light control method with edge computing and ambient edge intelligence, comprising a data center and an edge computing component. The data center component first uses a service discovery module to discover available devices that can serve as edge servers and converts the discovered devices into edge servers. After a camera captures a video frame, each frame is transmitted via two paths: one for local detection and the other for remote detection. The edge computing component performs initial object detection on the local device and extracts objects that cross the boundary. The data center component captures video using a surveillance camera, and the client sends the video frames to the corresponding edge device via an established WiFi connection. The edge computing component's edge server decodes data in a receiving queue and then sends the objects to the client. The data center client integrates the results of local and remote object detection to derive a final result, which is then used to control the traffic light.

Description

Technical Field

[0001] The present invention relates to the fields of Internet of Things technology and edge computing technology, and in particular to a traffic light control method with environmental edge intelligence based on edge computing. Background Art

[0002] In the 21st century, with its rapidly evolving economy and technology, innovative mobile applications and services such as augmented reality, federated learning, object detection, and edge caching are rapidly emerging. These applications significantly increase the demand for computing resources, typically provided by cloud servers. However, due to heavy backhaul loads and long distances, communication latency is high. For traffic control services, acceptable latency ranges from 10 milliseconds to 100 milliseconds.

[0003] Edge computing is considered a complement to cloud computing and tends to provide elastic resources at the edge of the network. To reduce communication latency, edge devices that are physically closer to the network than the cloud infrastructure are used. These edge devices are called edge servers. EdgeGo is a mobile resource sharing framework that uses mobile edge servers (mounted on mobile robots or drones) to achieve low-cost deployment of 6G edge computing. Unfortunately, the framework does not fully utilize the rich processing power of increasingly powerful environmental devices such as smartphones. MiVeCC is a vehicle-edge-cloud simulation framework designed to optimize vehicle cooperative control in large areas with multiple intersections. However, it does not consider resource budgets and the impact of communication and computation changes.

[0004] Instead of pre-deploying base stations and servers, the present invention offloads tile-level object detection to devices owned by the crowd to achieve intelligent traffic light control. The main work of the present invention is how to discover available devices and how to offload tasks. Among them, the access point periodically sends beacons embedded with the required services to discover available devices as edge servers. The detection model runs on the local device to perform initial object detection. Then objects that cross any boundaries are extracted. In order to improve detection accuracy and shorten decoding time, each frame is divided into multiple tiles for fine-grained detection on the selected server. An effective redundant server mechanism ensures the reliability of task completion. In addition, the CPU load of the server is used to estimate its computing power and resource payment is used to estimate its computing power. Finally, the objects detected locally and remotely are integrated to obtain the final result, which is the basis of a low-cost and reliable system. Summary of the Invention

[0005] To address the above-mentioned problems in the prior art, this application proposes a traffic light control method with edge computing and environmental edge intelligence, which is characterized by including a data center part and an edge computing part. The specific steps are as follows:

[0006] In step 1, the data center first periodically sends beacons embedded with the required services through access points. This beacon uses the service discovery module to identify available devices that can serve as edge servers. Once a device is found, it is converted into an edge server. Once ready, the camera begins capturing video frames, each of which is transmitted via two paths: one for local detection and the other for remote detection.

[0007] In step 2, the edge computing part first performs initial object detection by running the detection model on the local device, extracting objects that cross the boundary and performing video analysis within the local resource constraints;

[0008] In step 3, the data center client first crops each video frame into several main tiles. Each main tile is then filled with surrounding overlapping tiles to create a new sub-frame. This eliminates the dependencies between tiles, thus achieving tile-level parallelism.

[0009] Step 4: Based on step 3, the data center encodes each subframe and puts it into the sending queue, waiting to be sent to the selected edge server;

[0010] Step 5: For the video frame sent in step 4, the edge computing server decodes the data in the receiving queue and performs object detection. The server then sends the object (including subframe index, bounding box, label and score) to the client.

[0011] In step 6, the client in the data center integrates the results of the local and remote detected objects to obtain the final result, and controls the traffic light according to the final result.

[0012] Preferably, the data center part includes: a service manager, a video streaming media, a video transmitter, a result receiver and an object detector, and the edge computing part includes a service updater, a video decoder and an object detector of the edge node.

[0013] Preferably, the step 1 further comprises:

[0014] Step 11. The service discovery module is equipped with a connectionless protocol, which is a connectionless service discovery protocol based on WiFi beacons. It allows edge devices to know whether they have the service requested by the client even when they are disconnected. It requires the client to establish a communication connection (such as WiFi) with devices within the coverage area to discover available devices. Service information is embedded in the beacon. Nearby mobile devices perform service queries by parsing the beacon information. If the mobile device has the service, it will actively establish a WiFi connection with the client. Next, it will register its service information, including the service port, with the client, and then become a member of the server set. Each server must periodically report its device status, including IP, CPU load, and RSSI (received signal strength indicator).

[0015] Step 12: Processing a service query means checking whether an element with a given attribute value belongs to a given set. Bloom filters can be used as an optimization method.

[0016] Preferably, the Bloom filter optimization method is:

[0017] Step 121. Use multiple independent hash functions to map the elements to different positions / bits in the array and set the bit values ​​at the corresponding positions to 1. If any bit is 0, then the element is definitely not in the set. Otherwise, if the bit is 1 at all positions, then either the element is in the set or the bit happens to be set to 1 when another element is inserted (resulting in a false positive rate of F_r). The Fr of the Bloom filter can be optimized by the number of elements (n, the number of service instances on each server in this article), the number of hash functions (k), and the number of bits (b, the size of the bit array). Then the calculation formula for Fr is

[0018] Step 122, service discovery optimization: Each server has a Bloom filter. The client hashes the currently requested service name string into a location index based on the number of hash functions and embeds these bytes into the beacon. Each server parses the beacon after receiving it and queries each index through the Bloom filter. Only when each index in the server's Bloom filter bit array is 1 can the server provide the requested service and then join the client's wireless network. Otherwise, the server will not join the network. In addition, the query time complexity of the Bloom filter query is O(1) / O(k), while the time complexity of the traditional linear query is O(n). This method can effectively avoid the client communicating with each device separately. Especially when there are many devices in the environment, the communication cost can be greatly reduced.

[0019] Preferably, the step 2 further comprises:

[0020] Two different lightweight CNN models are used for object detection on the client and server. On the client, YOLOv5s is selected as the detection model. On the server, the lightweight Nanodet-Plus model is selected.

[0021] Preferably, the step 3 further comprises:

[0022] Step 31, determine the blocks containing small objects:

[0023] Use the maximum size of all small objects to determine the overlap size. Use the heavyweight detection model to detect the video and consider the detection results as ground truth. If the client object detector cannot classify a pixel patch into the correct category, but the heavyweight detection model can classify the pixel patch into the correct category, then the pixel patch is considered to contain a small object;

[0024] Step 32, tile-level partitioning:

[0025] Performing object detection on part of a frame may split a single object instance into multiple incorrect instances. Local detection is used to extract large and medium-sized objects, and long-range detection is used to extract small objects. A new method is used here to integrate the detection results. Three boundaries (white dashed lines) are uniformly annotated on a frame image, and then local detection is performed to obtain a group of objects. Objects that cross any boundaries are extracted. For long-range detection, the frame is first cropped into 6 major tiles / blocks based on boundaries and overlap size. Each major tile is padded with overlapping tiles to form a new subframe to alleviate the dependencies mentioned above. These subframes are then offloaded for long-range detection.

[0026] Preferably, the step 4 further comprises:

[0027] The sending module is designed to accelerate the detection pipeline. For each subtask / subframe, the cost of each server is estimated based on the payment and end-to-end (E2E) latency. The server with the lowest cost and meeting the deadline is then selected. If no server meets the deadline, the server with the lowest E2E latency is selected to ensure that the current subtask is not missed. After selecting the first optimal server for each subtask, the reliability of the current subtask determines whether to select additional servers, i.e., redundant servers. This offload control decision optimizes server selection, making the system more economical, reliable, accurate, and fast.

[0028] Preferably, the step 6 further comprises:

[0029] Step 61: Integrate the detection results of the client and server. All the detection results of the client and server are collected as initial results and then added to the detection list.

[0030] Step 62: Duplicate object detections may appear in this initial result. The initial result is consolidated based on labels, bounding box intersections, and class scores. Objects that do not cross any boundaries in the entire frame are removed from the detection list.

[0031] Step 63: If the intersection of the detection boxes between the tiles is higher than 75%, then the one with the higher score is considered a better choice and the other one is removed.

[0032] Step 64: If the intersection of the detection boxes between the full frame (at this point only objects crossing any boundaries are retained) and the tile exceeds 75%, the detection box in the full frame is considered a better choice and the other is removed.

[0033] The above technical features can be combined in various suitable ways or replaced by equivalent technical features, as long as the purpose of the present invention can be achieved.

[0034] The traffic light control method with edge computing and environmental edge intelligence provided by the present invention has at least the following advantages compared with the prior art:

[0035] First, this method effectively addresses the high capital and operating costs of base stations, including the purchase of physical equipment, site rental, installation and maintenance costs, and the additional costs of electricity and backhaul transmission. Only the resource contributions of volunteers are required. Second, this method also addresses the long deployment cycles typically required for base station deployment. The crowd-assisted system in this invention is highly flexible and can be rapidly expanded as needed. Finally, frequent disasters such as heavy rains, landslides, forest fires, and earthquakes can damage network infrastructure, resulting in the absence of E2E communication networks in the area. However, the opportunistic network provisioning assisted by the crowd in this invention ensures post-disaster availability. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The present invention will be described in more detail below based on embodiments and with reference to the accompanying drawings, wherein:

[0037] Figure 1 A flow chart of a traffic light control method with ambient edge intelligence based on edge computing of the present invention is shown; DETAILED DESCRIPTION

[0038] The present invention will be further described below with reference to the accompanying drawings.

[0039] Example 1

[0040] like Figure 1As shown, the present invention provides a traffic light control method with ambient edge intelligence based on edge computing, including a data center part and an edge computing part. The data center part includes: a service manager, a video streamer, a video transmitter, a result receiver, and an object detector, while the edge computing part includes a service updater, a video decoder, and an object detector of the edge node.

[0041] First, the data center needs to confirm whether there are available devices that can serve as edge servers. Edge servers, as key components in modern distributed computing architectures, enable rapid response and processing of data from data sources, particularly in scenarios requiring high real-time performance, such as intelligent traffic monitoring. If no edge servers are available, the process ends immediately.

[0042] If there is an available edge server and the conversion is successful, the camera starts capturing video frames. These captured video frames will be used for local detection and remote detection respectively.

[0043] The edge computing part first performs local detection. In the local path, the CNN model runs locally to perform preliminary object detection and extract objects that cross the boundary.

[0044] While local detection is ongoing, remote detection also begins simultaneously. The data center crops each video frame into several primary tiles. Each primary tile is then padded with surrounding overlapping tiles to create a new subframe. This eliminates dependencies between tiles, enabling tile-level detection parallelism.

[0045] Next, in the local data center, each subframe on the device is encoded into a specific file type. This process converts the raw video data into a format more suitable for transmission and storage, while retaining the necessary information for subsequent decoding and detection. Once encoded, these files are placed in a sending queue, awaiting delivery to the selected edge server.

[0046] After receiving data files from the data center, the edge computing part first decodes these files and restores them to the original sub-frame data. The decoded sub-frame data then undergoes the object detection phase, where the system analyzes each sub-frame, identifies the objects within, and classifies and locates them.

[0047] After object detection is complete, the edge computing component integrates the previous local detection results to generate a comprehensive object detection result, which is then sent to the client for use.

[0048] After receiving the object detection results from the edge server, some clients in the data center begin to integrate the results of local and remote detections and control related devices based on the final comprehensive results. In this invention, the final device is a traffic light.

[0049] The entire process is an iterative one: the camera continuously captures new video frames, the data center continuously processes, encodes, and sends new data files, and the edge computing component continuously decodes, detects, and synthesizes the new results. This system enables real-time processing and response to video streams, providing powerful technical support for fields such as intelligent transportation and security monitoring.

[0050] This method has significant advantages over previous technologies. First, this method can effectively solve the problem of expensive capital and operating costs of base stations, such as purchasing physical equipment, site leasing, installation and maintenance costs, additional costs for electricity and backhaul transmission, and only requires paying the resource contributions of volunteers. Second, this method also solves the problem that base station deployment usually requires a long deployment cycle. The crowd-assisted system in the present invention is very flexible and can be quickly expanded as needed. Finally, frequent disasters such as heavy rains, landslides, forest fires and earthquakes may damage network infrastructure, resulting in the absence of E2E communication networks in the area, but the opportunistic network supply assisted by the crowd in the present invention can ensure post-disaster availability.

[0051] Example 2

[0052] A traffic light control method with edge computing and environmental edge intelligence includes a data center part and an edge computing part, as follows:

[0053] In step 1, the data center first periodically sends beacons embedded with the required services through access points. This beacon uses the service discovery module to identify available devices that can serve as edge servers. Once a device is found, it is converted into an edge server. Once ready, the camera begins capturing video frames, each of which is transmitted via two paths: one for local detection and the other for remote detection.

[0054] In step 2, the edge computing part first performs initial object detection by running the detection model on the local device, extracting objects that cross the boundary and performing video analysis within the local resource constraints;

[0055] In step 3, the data center client first crops each video frame into several main tiles. Each main tile is then filled with surrounding overlapping tiles to create a new sub-frame. This eliminates the dependencies between tiles, thus achieving tile-level parallelism.

[0056] Step 4: Based on step 3, the data center encodes each subframe and puts it into the sending queue, waiting to be sent to the selected edge server;

[0057] Step 5: For the video frame sent in step 4, the edge computing server decodes the data in the receiving queue and performs object detection. The server then sends the object (including subframe index, bounding box, label and score) to the client.

[0058] In step 6, the client in the data center integrates the results of the local and remote detected objects to obtain the final result, and controls the traffic light according to the final result.

[0059] The data center part includes: a service manager, a video streamer, a video transmitter, a result receiver and an object detector, and the edge computing part includes a service updater, a video decoder and an object detector of the edge node.

[0060] The step 1 further comprises:

[0061] Step 11: Connectionless protocol of service discovery module:

[0062] This protocol is a connectionless service discovery protocol based on WiFi beacons, which allows edge devices to know whether they have the service requested by the client even when disconnected. It requires the client to establish a communication connection (such as WiFi) with devices within the coverage area to discover available devices. Service information is embedded in the beacon. Nearby mobile devices perform service queries by parsing the beacon information. If the mobile device has the service, it will actively establish a WiFi connection with the client. Next, it will register its service information, including the service port, with the client, and then become a member of the server set. Each server must periodically report its device status, including IP, CPU load, and RSSI (Received Signal Strength Indicator).

[0063] Step 12: Optimize Bloom filter:

[0064] Processing a service query means checking whether an element with a given property value belongs to a given set. Bloom filters can be used as an optimization method.

[0065] The Bloom filter optimization method is:

[0066] Step 121, Standard Bloom filter:

[0067] Use multiple independent hash functions to map elements to different positions / bits in the array and set the bit value of the corresponding position to 1. If any bit is 0, then the element is definitely not in the set. Otherwise, if the bit is 1 in all positions, then either the element is in the set or the bit happened to be set to 1 when another element was inserted (resulting in a false positive rate of F_r). The Fr of the Bloom filter can be optimized by the number of elements (n, the number of service instances on each server in this article), the number of hash functions (k), and the number of bits (b, the size of the bit array). Then the calculation formula for Fr is

[0068] Step 122: Service discovery optimization:

[0069] Each server has a Bloom filter. The client hashes the currently requested service name string into a location index based on the number of hash functions and embeds these bytes in the beacon. Each server parses the beacon after receiving it and queries each index through the Bloom filter. Only when each index in the server's Bloom filter bit array is 1 can the server provide the requested service and then join the client's wireless network. Otherwise, the server will not join the network. In addition, the query time complexity of the Bloom filter query is O(1) / O(k), while the time complexity of the traditional linear query is O(n). This method can effectively avoid the client communicating with each device separately. Especially when there are many devices in the environment, the communication cost can be greatly reduced.

[0070] The step 2 further comprises:

[0071] Two different lightweight CNN models are used for object detection on the client and server. On the client, YOLOv5s is selected as the detection model. On the server, the lightweight Nanodet-Plus model is selected.

[0072] The step 3 further comprises:

[0073] Step 31, determine the blocks containing small objects:

[0074] Use the maximum size of all small objects to determine the overlap size. Use the heavyweight detection model to detect the video and consider the detection results as ground truth. If the client object detector cannot classify a pixel patch into the correct category, but the heavyweight detection model can classify the pixel patch into the correct category, then the pixel patch is considered to contain a small object;

[0075] Step 32, tile-level partitioning:

[0076] Performing object detection on part of a frame may split a single object instance into multiple incorrect instances. Local detection is used to extract large and medium-sized objects, and long-range detection is used to extract small objects. A new method is used here to integrate the detection results. Three boundaries (white dashed lines) are uniformly annotated on a frame image, and then local detection is performed to obtain a group of objects. Objects that cross any boundaries are extracted. For long-range detection, the frame is first cropped into 6 major tiles / blocks based on boundaries and overlap size. Each major tile is padded with overlapping tiles to form a new subframe to alleviate the dependencies mentioned above. These subframes are then offloaded for long-range detection.

[0077] The step 4 further comprises:

[0078] The sending module is designed to accelerate the detection pipeline. For each subtask / subframe, the cost of each server is estimated based on payment and end-to-end (E2E) latency. The server with the lowest cost and meeting the deadline is then selected. If no server meets the deadline, the server with the lowest E2E latency is selected to ensure that the current subtask is not missed. After selecting the first optimal server for each subtask, the reliability of the current subtask determines whether to select additional servers, i.e., redundant servers. This offload control decision optimizes server selection, making the system more economical, reliable, accurate, and fast.

[0079] The step 6 further comprises:

[0080] Step 61: Integrate the detection results of the client and server. All the detection results of the client and server are collected as initial results and then added to the detection list.

[0081] Step 62: Duplicate object detections may appear in this initial result. The initial result is consolidated based on labels, bounding box intersections, and class scores. Objects that do not cross any boundaries in the entire frame are removed from the detection list.

[0082] Step 63: If the intersection of the detection boxes between the tiles is higher than 75%, then the one with the higher score is considered a better choice and the other one is removed.

[0083] Step 64: If the intersection of the detection boxes between the full frame (at this point only objects crossing any boundaries are retained) and the tile exceeds 75%, the detection box in the full frame is considered a better choice and the other is removed.

[0084] Example 3

[0085] Using a connectionless protocol in the service controller. This paper proposes a connectionless service discovery protocol based on WiFi beacons, which allows edge devices to determine whether they possess the service requested by the client even when disconnected. Due to limited computing resources, edge servers can only store or execute a limited number of service instances. This protocol requires the client to establish a communication connection (such as WiFi) with devices within range to discover available devices. In densely populated areas, the number of devices around is enormous. If the client frequently communicates with all devices, even with some that lack services, this wastes energy and time. Therefore, an effective mechanism must be designed to allow the client to discover target devices that possess the required service. Because the WiFi protocol has a characteristic that, even when no data is being transmitted, it must periodically (typically 100 milliseconds) broadcast beacons to announce the existence of the network, allowing nearby devices to join the network. This characteristic can be exploited to embed service information in the beacons. Nearby mobile devices query for services by parsing the beacon information. The design is that if a mobile device possesses a service, it proactively establishes a WiFi connection with the client. Next, it registers its service information, including the service port, with the client and becomes a member of the server set. Each server must periodically report its device status, including IP, CPU load, and RSSI (Received Signal Strength Indicator). There is almost no additional communication overhead in the service discovery process.

[0086] There are two optimization methods for using Bloom filters in service controllers: the first is the standard Bloom filter and the second is service discovery optimization.

[0087] The first type, standard Bloom filter:

[0088] Use multiple independent hash functions to map elements to different positions / bits in the array and set the bit value of the corresponding position to 1. If any bit is 0, then the element is definitely not in the set. Otherwise, if the bit is 1 in all positions, then either the element is in the set or the bit happened to be set to 1 when another element was inserted (resulting in a false positive rate of F_r). The Fr of the Bloom filter can be optimized by the number of elements (n, the number of service instances on each server in this article), the number of hash functions (k), and the number of bits (b, the size of the bit array). Then the calculation formula for F_r is

[0089] Second, service discovery optimization:

[0090] Each server has a Bloom filter. The client hashes the currently requested service name string into a location index based on the number of hash functions and embeds these bytes in the beacon. Each server parses the beacon after receiving it and queries each index through the Bloom filter. Only when each index in the server's Bloom filter bit array is 1 can the server provide the requested service and then join the client's wireless network. Otherwise, the server will not join the network. In addition, the query time complexity of the Bloom filter query is O(1) / O(k), while the time complexity of the traditional linear query is O(n). This method can effectively avoid the client communicating with each device separately. Especially when there are many devices in the environment, the communication cost can be greatly reduced.

[0091] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the invention. It should be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in ways other than those described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be employed in conjunction with other described embodiments.

Claims

1. A traffic light control method with edge computing and environmental edge intelligence, characterized in that: Including the data center part and the edge computing part, the specific steps are as follows: In step 1, the data center periodically sends beacons embedded with the required services through access points. The service discovery module then discovers devices that can serve as edge servers. Once a device is found, it is converted to an edge server. The camera begins capturing video frames, each of which is transmitted via two paths: one for local detection and the other for remote detection. In step 2, the edge computing part first runs the detection model on the local device to perform initial object detection, extract objects that cross the boundary, and perform video analysis within the local resource constraints; In step 3, the data center part crops each video frame into several main tiles. Then, each main tile is filled with surrounding overlapping tiles to create a new sub-frame. By eliminating the dependencies between tiles, tile-level detection parallelism is achieved. Step 4: Based on step 3, the data center encodes each subframe and puts it into the sending queue, waiting to be sent to the selected edge server; Step 5: For the video frame sent in step 4, the edge computing part decodes the data in the receiving queue and performs object detection; The server then sends the object to the client; In step 6, the data center integrates the local and remote detection results to obtain the final result and controls the traffic lights according to the final result.

2. The traffic light control method with edge computing and environmental edge intelligence according to claim 1 is characterized in that: The data center part includes: a service manager, a video streamer, a video transmitter, a result receiver and an object detector, and the edge computing part includes a service updater, a video decoder and an object detector of the edge node.

3. The traffic light control method with edge computing and environmental edge intelligence according to claim 1, characterized in that: The service discovery module uses a connectionless service discovery protocol based on WiFi beacons, allowing edge devices to know whether they have the service requested by the client even when disconnected. The client establishes a communication connection with devices within the coverage area to discover available devices, and embeds service information in the beacon. Nearby mobile devices query services by parsing beacon information; if the mobile device has the service, it will actively establish a WiFi connection with the client; next, the mobile device will register its service information with the client and become a member of the server set.

4. The traffic light control method with edge computing and environmental edge intelligence according to claim 1, characterized in that: Optimize service discovery. The optimization methods for service discovery are: Use multiple independent hash functions to map elements to different positions in the array and set the bit value of the corresponding position to 1; the false positive rate Fr of the filter can be optimized by the number of elements n, the number of hash functions k and the number of bits b; the calculation formula of the false positive rate Fr is 5. The traffic light control method with edge computing and environmental edge intelligence according to claim 4 is characterized in that: Each server has filters; the client hashes the currently requested service name string into a location index according to the number of hash functions and embeds these bytes in the beacon; each server parses the beacon after receiving it and queries each index through the filter; only when each index in the server's filter bit array is 1 can the server provide the requested service and then join the client's wireless network; otherwise, the server will not join the network.

6. The traffic light control method with edge computing and environmental edge intelligence according to claim 1, characterized in that: Two different lightweight CNN models are used on the client and server for object detection.

7. The traffic light control method with edge computing and environmental edge intelligence according to claim 1, characterized in that: The step 3 further comprises: Step 31, using the maximum size of the small object to determine the overlap size; using the heavyweight detection model to detect the video, and treating the detection results as ground truth; if the client object detector cannot classify the pixel block into the correct category, but the heavyweight detection model can classify the pixel block into the correct category, then the pixel block is considered to contain a small object; In step 32, local detection is used to extract large and medium-sized objects, and remote detection is used to extract small objects; multiple boundaries are uniformly marked on a frame image, and then local detection is performed; the frame is cropped into multiple main tiles / blocks based on the boundaries and overlap size; each main tile is filled with overlapping tiles to form a new subframe, and then these subframes are unloaded for remote detection.

8. The traffic light control method with edge computing and environmental edge intelligence according to claim 1, characterized in that: The step 4 further comprises: The data center part includes a sending module, which is designed to accelerate the detection pipeline. For each subtask / subframe, the cost of each server is estimated based on the payment and E2E latency. Then, the server with the lowest cost and meeting the deadline requirements is selected. If no server meets the deadline requirements, the server with the smallest E2E latency is selected to ensure that the current subtask is not missed. After selecting the first optimal server for each subtask, the decision on whether to continue selecting more servers is made based on the reliability of the current subtask.

9. The traffic light control method with edge computing and environmental edge intelligence according to claim 1, characterized in that: The step 6 further comprises: Step 61: The detection results of the client and the server are collected as initial results and then added to the detection list; Step 62: Integrate the initial results based on labels, bounding box intersections, and category scores; objects that do not cross any boundaries in the entire frame are removed from the detection list; Step 63: If the intersection of the detection boxes between the tiles is higher than 75%, then the one with the higher score will be considered a better choice and the other one will be removed. Step 64: If the intersection of the detection boxes between the full frame and the tile exceeds 75%, the detection box in the full frame is considered a better choice and the other one is removed.

Images