Obstacle avoidance method and system for intelligent inspection vehicle of farm

Abstract

The invention discloses an obstacle avoidance method and system for a farm intelligent inspection vehicle, and belongs to the technical field of automatic obstacle avoidance, and the method comprises the steps: obtaining a key position of a farm, and constructing a farm map structure based on the key position; acquiring the position of an obstacle around the inspection vehicle through a sensor of the inspection vehicle, and feeding back the position of the obstacle to the weight of the edge of the graph structure; based on a Dijkstra algorithm, calculating the shortest running path of the inspection vehicle for the graph structure; constructing a depth Q network model through the number of nodes of the graph structure, and updating the value of the shortest path distance from the current node to the reachable node in the depth Q network model according to the shortest path of the inspection vehicle to complete real-time updating of the depth Q network model; and the inspection vehicle inquires the updated depth Q network model in real time in the inspection process to dynamically adjust the advancing route. According to the invention, the optimal driving route can be quickly calculated, and the total driving distance and time of the inspection vehicle are reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of automatic obstacle avoidance, and in particular relates to an obstacle avoidance method and system for an intelligent inspection vehicle in a farm. Background Art

[0002] With the rapid development of large-scale caged poultry farming technology and indoor agricultural technology and equipment, indoor mobile robots have emerged. Mobile robots have a wide range of applications in farms because of their flexible movements, strong adaptability to the ground environment, low cost, and ability to obtain overall information and local specific information in the farm. Due to the complex environment in the farm, mobile robots will inevitably encounter various types of obstacles when operating, including obstacles of different sizes, static or dynamic. Static obstacles include cage columns, cargo boxes, feeders, etc.; dynamic obstacles include all people or objects with movement speed and acceleration, such as staff in the house. In order to improve efficiency and ensure the safety of personnel and materials, mobile robot obstacle avoidance has a high practical significance.

[0003] At present, mobile robots used in farms mainly use passive stop avoidance based on a single active sensor, that is, using ultrasonic sensors, laser radars, infrared sensors, etc. to actively send signals and receive sensor data of return signals. When the data reaches a certain threshold, it immediately stops moving and waits for the obstacle to disappear or switches to human guidance. Due to the disadvantage of incomplete information obtained by sensors, flexibility is low and operating efficiency is also greatly affected. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes an obstacle avoidance method and system for an intelligent inspection vehicle in a farm to solve the problems existing in the above prior art.

[0005] To achieve the above object, the present invention provides an obstacle avoidance method for an intelligent inspection vehicle in a farm, comprising:

[0006] Acquire key locations of the farm, and construct a farm graph structure based on the key locations;

[0007] Obtaining the positions of obstacles around the inspection vehicle through the sensors of the inspection vehicle, and feeding the obstacle positions back into the weights of the edges of the graph structure;

[0008] Calculate the shortest path for the inspection vehicle to run based on the graph structure based on the Dijkstra algorithm;

[0009] A deep Q network model is constructed by the number of nodes in the graph structure, and the value of the shortest path distance from the current node to the reachable node in the deep Q network model is updated according to the shortest path of the inspection vehicle, so as to complete the real-time update of the deep Q network model;

[0010] During the inspection process, the inspection vehicle queries the updated deep Q network model in real time to dynamically adjust the route.

[0011] Preferably, the process of obtaining the key positions of the farm map and constructing the farm map structure based on the key positions includes:

[0012] Defining key locations of the farm, and defining the key locations as nodes of a graph;

[0013] Based on the actual reachability and connectivity, edges are defined between the nodes of the graph to construct the farm graph structure.

[0014] Preferably, the process of obtaining the obstacle positions around the inspection vehicle through the sensors of the inspection vehicle and feeding the obstacle positions back to the weights of the edges of the graph structure includes:

[0015] Installing sensors on the inspection vehicle, and acquiring surrounding environment data of the inspection vehicle based on the sensors;

[0016] Identify the surrounding environment data by machine vision method to obtain surrounding obstacle position data;

[0017] The detected surrounding obstacle position data is converted into changes in weights or edges in the graph and fed back to the graph structure for modification.

[0018] Preferably, the process of calculating the shortest path for the inspection vehicle to run based on the graph structure using the Dijkstra algorithm includes:

[0019] Taking a node in the graph structure as the starting point, initialize the distance array and record the shortest distances from the starting point to all other nodes. At the beginning, except for the distance from the starting point to itself being 0, the shortest distances from the starting point to all other nodes are set to infinity.

[0020] Starting from the starting point, all unprocessed nodes are explored step by step, the node with the shortest distance is selected, this path is confirmed as the shortest path for the inspection vehicle to run, and the shortest path distance of its adjacent nodes is updated.

[0021] Preferably, the process of constructing a deep Q network model by the number of nodes in the graph structure includes:

[0022] Constructing an action space available to the inspection vehicle based on the number of nodes, constructing a Q network model based on the fully connected neural network and the action space available to the inspection vehicle, taking the inspection vehicle state representation as the input of the Q network model, and outputting a Q value for each action;

[0023] Constructing a loss function of the Q network model, wherein the loss function is an objective function in a Q-learning algorithm;

[0024] The Q network model is trained based on experience replay and the target network. The Q network model learns the optimal action strategy through interaction with the environment to maximize the long-term cumulative reward, thereby generating the deep Q network model.

[0025] Preferably, the process of completing the real-time update of the deep Q network model includes:

[0026] The shortest path for the inspection vehicle to run is calculated by the Dijkstra algorithm on the graph structure as the negative value of the immediate reward to update the corresponding parameters in the deep Q network model, so as to minimize the cost of reaching the obstacle avoidance state through the inspection vehicle driving action in the deep Q network model.

[0027] Preferably, the process of the inspection vehicle dynamically adjusting the route by querying the updated deep Q network model in real time during the inspection process includes:

[0028] The inspection vehicle obtains surrounding environment information in real time;

[0029] The surrounding environment information is converted into morphological features and then input into the deep Q network model to generate a local guidance path;

[0030] The local guidance path is repeatedly obtained to generate a planned route, and the inspection vehicle's travel route is adjusted based on the planned route.

[0031] Preferably, the process of adjusting the route of the inspection vehicle based on the planned route further includes:

[0032] If there is a difference between the morphological features and the key nodes of the graph structure, it is judged that a new obstacle has appeared or the original path is no longer feasible, and the Dijkstra algorithm is rerun to update the affected part of the path, and the deep Q network model is adjusted accordingly.

[0033] To achieve the above object, the present invention also provides an obstacle avoidance system for a farm intelligent inspection vehicle, comprising:

[0034] A construction subsystem is used to obtain key locations of the farm and construct a farm graph structure based on the key locations;

[0035] A feedback subsystem, used to obtain the position of obstacles around the inspection vehicle through the sensors of the inspection vehicle, and feed the obstacle position back to the weight of the edge of the graph structure;

[0036] A calculation subsystem, used for calculating the shortest path for the inspection vehicle to run on the graph structure based on the Dijkstra algorithm;

[0037] An updating subsystem is used to construct a deep Q network model according to the number of nodes in the graph structure, and to update the value of the shortest path distance from the current node to the reachable node in the deep Q network model according to the shortest path of the inspection vehicle, so as to complete the real-time update of the deep Q network model;

[0038] The adjustment subsystem is used for the inspection vehicle to query the updated deep Q network model in real time during the inspection process and dynamically adjust the route.

[0039] Compared with the prior art, the present invention has the following advantages and technical effects:

[0040] The Dijkstra algorithm in the present invention provides an effective solution for finding the shortest path from the starting point to the end point. In this way, the optimal obstacle avoidance route can be quickly calculated to reduce the total distance and time traveled by the inspection vehicle. Combined with the deep Q network, real-time adjustment of the obstacle avoidance route can be achieved in a dynamic environment, allowing the inspection vehicle to flexibly respond to emergencies, such as the movement of animals or the emergence of new obstacles. The deep Q network can recognize complex patterns and relationships by learning from a large amount of historical data. The deep Q network is used to handle scenarios that require complex decision-making and predictions. In the ever-changing farm environment, this can significantly improve the autonomy and efficiency of the inspection vehicle's obstacle avoidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0042] Figure 1 This is a flow chart of an obstacle avoidance method for a farm smart inspection vehicle according to an embodiment of the present invention;

[0043] Figure 2 Schematic diagram of the obstacle avoidance system of the smart inspection vehicle for farms according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0045] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0046] Embodiment 1

[0047] like Figure 1As shown, this embodiment provides an obstacle avoidance method for a farm intelligent inspection vehicle, including:

[0048] Acquire key locations of the farm, and construct a farm graph structure based on the key locations;

[0049] Obtaining the positions of obstacles around the inspection vehicle through the sensors of the inspection vehicle, and feeding the obstacle positions back into the weights of the edges of the graph structure;

[0050] Calculate the shortest path for the inspection vehicle to run based on the graph structure based on the Dijkstra algorithm;

[0051] A deep Q network model is constructed by the number of nodes in the graph structure, and the value of the shortest path distance from the current node to the reachable node in the deep Q network model is updated according to the shortest path of the inspection vehicle, so as to complete the real-time update of the deep Q network model;

[0052] During the inspection process, the inspection vehicle queries the updated deep Q network model in real time to dynamically adjust the route.

[0053] Further optimizing the scheme, the process of obtaining the key positions of the farm map and constructing the farm map structure based on the key positions includes:

[0054] Defining key locations of the farm, and defining the key locations as nodes of a graph;

[0055] Based on the actual reachability and connectivity, edges are defined between the nodes of the graph to construct the farm graph structure.

[0056] Further optimizing the solution, the process of obtaining the obstacle positions around the inspection vehicle through the sensors of the inspection vehicle and feeding the obstacle positions back to the weights of the edges of the graph structure includes:

[0057] Installing sensors on the inspection vehicle, and acquiring surrounding environment data of the inspection vehicle based on the sensors;

[0058] Identify the surrounding environment data by machine vision method to obtain surrounding obstacle position data;

[0059] The detected surrounding obstacle position data is converted into changes in weights or edges in the graph and fed back to the graph structure for modification.

[0060] To further optimize the solution, the process of calculating the shortest path for the inspection vehicle based on the graph structure using the Dijkstra algorithm includes:

[0061] Taking a node in the graph structure as the starting point, initialize the distance array and record the shortest distances from the starting point to all other nodes. At the beginning, except for the distance from the starting point to itself being 0, the shortest distances from the starting point to all other nodes are set to infinity.

[0062] Starting from the starting point, all unprocessed nodes are explored step by step, the node with the shortest distance is selected, this path is confirmed as the shortest path for the inspection vehicle to run, and the shortest path distance of its adjacent nodes is updated.

[0063] Further optimizing the scheme, the process of constructing the deep Q network model by the number of nodes of the graph structure includes:

[0064] Constructing an action space available to the inspection vehicle based on the number of nodes, constructing a Q network model based on the fully connected neural network and the action space available to the inspection vehicle, taking the inspection vehicle state representation as the input of the Q network model, and outputting a Q value for each action;

[0065] Constructing a loss function of the Q network model, wherein the loss function is an objective function in a Q-learning algorithm;

[0066] The Q network model is trained based on experience replay and the target network. The Q network model learns the optimal action strategy through interaction with the environment to maximize the long-term cumulative reward, thereby generating the deep Q network model.

[0067] Further optimizing the scheme, the process of completing the real-time update of the deep Q network model includes:

[0068] The shortest path for the inspection vehicle to run is calculated by the Dijkstra algorithm on the graph structure as the negative value of the immediate reward to update the corresponding parameters in the deep Q network model, so as to minimize the cost of reaching the obstacle avoidance state through the inspection vehicle driving action in the deep Q network model.

[0069] To further optimize the solution, the inspection vehicle queries the updated deep Q network model in real time during the inspection process to dynamically adjust the route, including:

[0070] The inspection vehicle obtains surrounding environment information in real time;

[0071] The surrounding environment information is converted into morphological features and then input into the deep Q network model to generate a local guidance path;

[0072] The local guidance path is repeatedly obtained to generate a planned route, and the inspection vehicle's travel route is adjusted based on the planned route.

[0073] To further optimize the solution, the process of adjusting the route of the inspection vehicle based on the planned route also includes:

[0074] If there is a difference between the morphological features and the key nodes of the graph structure, it is judged that a new obstacle has appeared or the original path is no longer feasible, and the Dijkstra algorithm is rerun to update the affected part of the path, and the deep Q network model is adjusted accordingly.

[0075] Embodiment 2

[0076] This embodiment provides an obstacle avoidance method for a farm intelligent inspection vehicle, including:

[0077] Each important location of the farm is defined as a node in the graph. These locations can include feeding areas, animal habitats, warehouses, entrances and exits, etc. At the same time, the current location of the smart inspection vehicle can also be set as a node.

[0078] The connections between nodes (i.e. the edges of the graph) are based on actual traversable paths. For example, a direct route between two nodes can define an edge. The weight of each edge is set based on the length of the path or the number of obstacles. For example, the longer the path or the more obstacles there are, the greater the weight.

[0079] Based on the Dijkstra algorithm, the shortest path for the inspection vehicle to run is calculated for the graph structure; the distance array is initialized to record the shortest distance from the starting point to all other nodes. At the beginning, except for the distance from the starting point to itself, which is 0, all others are set to infinity. Starting from the starting point, all unprocessed nodes are gradually explored, the node with the shortest distance is selected, this path is confirmed to be the shortest path, and the shortest path distance of its adjacent nodes is updated.

[0080] A deep Q network model is constructed by the number of nodes in the graph structure, and the value of the shortest path distance from the current node to the reachable node in the deep Q network model is updated according to the shortest path of the inspection vehicle, so as to complete the real-time update of the deep Q network model;

[0081] Deep Q-Network (DQN) is a technique that combines deep learning and Q-learning to solve decision-making problems with high-dimensional observation spaces. When building a deep Q-network model, the Q matrix is ​​usually not directly operated, but a deep neural network is used to approximate the Q function, which estimates the expected return of taking various actions in a given state.

[0082] First, we need to define the reinforcement learning environment, which includes the state space and the action space. The state should provide enough information for the network to make a decision, while the action space defines the possible actions to be taken in each state.

[0083] Designing the network architecture:

[0084] Input layer: The size of the input layer should match the dimension of the state space.

[0085] Hidden layers: Multiple hidden layers can be used, often including fully connected layers or convolutional layers with ReLU activation functions (especially when processing visual data).

[0086] Output layer: The size of the output layer should match the size of the action space. The output of each neuron represents the Q value of taking a specific action given the input state.

[0087] During training, a ∈-greedy strategy is usually used to balance exploration and exploitation, where ∈ is gradually reduced to make more use of the learned strategy.

[0088] For each state-action pair, the target Q-value is the immediate reward plus a discount factor times the maximum Q-value of the next state (from the network’s prediction).

[0089] Compute the shortest paths from the current state to all other states using Dijkstra's algorithm.

[0090] This information can be used as a heuristic to guide the DQN’s exploration strategy to prioritize exploring paths that are likely to be closer to the goal state.

[0091] The shortest path information obtained from Dijkstra's algorithm is used to make an initial estimate of the Q value in DQN.

[0092] For example, if the path cost from the cruising state to the stopped state of a cruise car is known, this cost can be used as the initial Q value for performing the corresponding action.

[0093] When performing experience replay, the transfer samples along the shortest path determined by the Dijkstra algorithm can be preferentially selected.

[0094] This prioritized sampling can help the network learn an effective strategy to reach the goal state faster.

[0095] In the early stages, the ∈-greedy strategy can be designed to prefer actions suggested by Dijkstra’s algorithm to quickly verify the effectiveness of these actions.

[0096] Over time, as the network learns more and more about the environment, it relies less on Dijkstra's algorithm and more on the output of the Q-network model.

[0097] Constructing a loss function of the Q network model, wherein the loss function is an objective function in a Q-learning algorithm;

[0098] For each state (s) and action (a) in Q learning (the action here can be understood as going from the current node to an adjacent node), update the Q value.

[0099] The Q value can be updated using the shortest path cost obtained by Dijkstra's algorithm. For example, if state (s') can be reached from state (s) via action (a), and this is one of the shortest paths, then the Q value update can be: [Q(s,a)=r(s,a,s')+\gamma\min_{a'}Q(s',a')] where (r(s,a,s')) is the immediate reward of reaching (s') from state (s) via action (a), which can usually be set to (-\text{edge weight}) or (-\text{path cost}), and (\gamma) is a discount factor.

[0100] The Q network model is trained based on experience replay and the target network. The Q network model learns the optimal action strategy through interaction with the environment to maximize the long-term cumulative reward, thereby generating the deep Q network model.

[0101] To further optimize the solution, the inspection vehicle queries the updated deep Q network model in real time during the inspection process to dynamically adjust the route, including:

[0102] The inspection vehicle obtains surrounding environment information in real time;

[0103] The surrounding environment information is converted into morphological features and then input into the deep Q network model to generate a local guidance path;

[0104] The local guidance path is repeatedly obtained to generate a planned route, and the inspection vehicle's travel route is adjusted based on the planned route.

[0105] To further optimize the solution, the process of adjusting the route of the inspection vehicle based on the planned route also includes:

[0106] If there is a difference between the morphological features and the key nodes of the graph structure, it is judged that a new obstacle has appeared or the original path is no longer feasible, and the Dijkstra algorithm is rerun to update the affected part of the path, and the deep Q network model is adjusted accordingly.

[0107] Embodiment 3

[0108] like Figure 2 As shown, this embodiment provides an obstacle avoidance system for a farm intelligent inspection vehicle, including:

[0109] A construction subsystem is used to obtain key locations of the farm and construct a farm graph structure based on the key locations;

[0110] A feedback subsystem, used to obtain the position of obstacles around the inspection vehicle through the sensors of the inspection vehicle, and feed the obstacle position back to the weight of the edge of the graph structure;

[0111] A calculation subsystem, used for calculating the shortest path for the inspection vehicle to run on the graph structure based on the Dijkstra algorithm;

[0112] An updating subsystem is used to construct a deep Q network model according to the number of nodes in the graph structure, and to update the value of the shortest path distance from the current node to the reachable node in the deep Q network model according to the shortest path of the inspection vehicle, so as to complete the real-time update of the deep Q network model;

[0113] The adjustment subsystem is used for the inspection vehicle to query the updated deep Q network model in real time during the inspection process and dynamically adjust the route.

[0114] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An obstacle avoidance method for an intelligent inspection vehicle in a farm, characterized in that: The following steps are involved: Acquire key locations of the farm, and construct a farm graph structure based on the key locations; Obtaining the positions of obstacles around the inspection vehicle through the sensors of the inspection vehicle, and feeding the obstacle positions back into the weights of the edges of the graph structure; Calculate the shortest path for the inspection vehicle to run based on the graph structure based on the Dijkstra algorithm; A deep Q network model is constructed by the number of nodes in the graph structure, and the value of the shortest path distance from the current node to the reachable node in the deep Q network model is updated according to the shortest path of the inspection vehicle, so as to complete the real-time update of the deep Q network model; During the inspection process, the inspection vehicle queries the updated deep Q network model in real time to dynamically adjust the route.

2. The obstacle avoidance method for the farm intelligent inspection vehicle according to claim 1, characterized in that: The process of obtaining the key positions of the farm map and constructing the farm map structure based on the key positions includes: Defining key locations of the farm, and defining the key locations as nodes of a graph; Based on the actual reachability and connectivity, edges are defined between the nodes of the graph to construct the farm graph structure.

3. The obstacle avoidance method for the farm intelligent inspection vehicle according to claim 1 is characterized in that: The process of obtaining the obstacle positions around the inspection vehicle through the sensors of the inspection vehicle and feeding the obstacle positions back to the weights of the edges of the graph structure includes: Installing sensors on the inspection vehicle and acquiring surrounding environment data of the inspection vehicle based on the sensors; Identify the surrounding environment data by machine vision method to obtain surrounding obstacle position data; The detected data of the positions of the surrounding obstacles are converted into changes in the weights or edges in the graph and fed back to the graph structure for modification.

4. The obstacle avoidance method for the farm intelligent inspection vehicle according to claim 1, characterized in that: The process of calculating the shortest path for the inspection vehicle based on the graph structure based on the Dijkstra algorithm includes: Taking a node in the graph structure as the starting point, initialize the distance array and record the shortest distances from the starting point to all other nodes. At the beginning, except for the distance from the starting point to itself being 0, the shortest distances from the starting point to all other nodes are set to infinity. Starting from the starting point, all unprocessed nodes are explored step by step, the node with the shortest distance is selected, this path is confirmed as the shortest path for the inspection vehicle to run, and the shortest path distance of its adjacent nodes is updated.

5. The obstacle avoidance method for the farm intelligent inspection vehicle according to claim 1, characterized in that: The process of constructing a deep Q network model by using the number of nodes in the graph structure includes: Constructing an action space available to the inspection vehicle based on the number of nodes, constructing a Q network model based on the fully connected neural network and the action space available to the inspection vehicle, taking the inspection vehicle state representation as the input of the Q network model, and outputting a Q value for each action; Constructing a loss function of the Q network model, wherein the loss function is an objective function in a Q-learning algorithm; The Q network model is trained based on experience replay and the target network. The Q network model learns the optimal action strategy through interaction with the environment to maximize the long-term cumulative reward, thereby generating the deep Q network model.

6. The obstacle avoidance method for the farm intelligent inspection vehicle according to claim 4, characterized in that: The process of completing the real-time update of the deep Q network model includes: The shortest path for the inspection vehicle to run is calculated by the Dijkstra algorithm on the graph structure as the negative value of the immediate reward to update the corresponding parameters in the deep Q network model, so as to minimize the cost of reaching the obstacle avoidance state through the inspection vehicle driving action in the deep Q network model.

7. The obstacle avoidance method for the farm intelligent inspection vehicle according to claim 1, characterized in that: The process by which the inspection vehicle dynamically adjusts its route by querying the updated deep Q network model in real time during the inspection process includes: The inspection vehicle obtains surrounding environment information in real time; The surrounding environment information is converted into morphological features and then input into the deep Q network model to generate a local guidance path; The local guidance path is repeatedly obtained to generate a planned route, and the inspection vehicle's travel route is adjusted based on the planned route.

8. The obstacle avoidance method for the farm intelligent inspection vehicle according to claim 7, characterized in that: The process of adjusting the route of the inspection vehicle based on the planned route also includes: If there is a difference between the morphological features and the key nodes of the graph structure, it is judged that a new obstacle has appeared or the original path is no longer feasible, and the Dijkstra algorithm is rerun to update the affected part of the path, and the deep Q network model is adjusted accordingly.

9. An obstacle avoidance system for a farm intelligent inspection vehicle, characterized in that: include: A construction subsystem is used to obtain key locations of the farm and construct a farm graph structure based on the key locations; A feedback subsystem, used to obtain the position of obstacles around the inspection vehicle through the sensors of the inspection vehicle, and feed the obstacle position back to the weight of the edge of the graph structure; A calculation subsystem, used for calculating the shortest path for the inspection vehicle to run on the graph structure based on the Dijkstra algorithm; An updating subsystem is used to construct a deep Q network model according to the number of nodes in the graph structure, and to update the value of the shortest path distance from the current node to the reachable node in the deep Q network model according to the shortest path of the inspection vehicle, so as to complete the real-time update of the deep Q network model; The adjustment subsystem is used for the inspection vehicle to query the updated deep Q network model in real time during the inspection process and dynamically adjust the route.

Images