Obstacle detouring judgment method and system for sanitation vehicle edge sweeping scene

Abstract

The invention discloses an obstacle detouring judgment method and system for a sanitation vehicle edge sweeping scene. According to the method disclosed by the invention, under the condition of only depending on a lightweight map (road boundary, intersection polygon and signal lamp associated information) and structured perception information (obstacle position, speed, acceleration, course, short-term prediction track, taillight / signal lamp state and optional visual semantics), each obstacle is judged to be 'Bypass' or 'Non-Bypass'; and the confidence coefficient is output for downstream bypass planning.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to an obstacle avoidance determination method and system for sanitation vehicles sweeping along the edges of roads. Background Technology

[0002] When sanitation vehicles perform edge cleaning on urban roads, they often need to drive close to the right edge of the road. The roadside frequently presents various dynamic and static obstacles, including parked vehicles, vehicles stopping to pick up goods, vehicles making short stops, vehicles queuing, two-wheeled vehicles, and pedestrians. Existing methods typically have the following shortcomings:

[0003] (1) Most of them are based on simple hard rules (if-else) or single threshold coverage, lacking modeling of observation uncertainty, and are prone to misjudgment under perceived noise.

[0004] (2) Scene recognition often adopts a hard coverage strategy (the scene is covered as soon as it is generated), which leads to false coverage and efficiency loss.

[0005] (3) The treatment of two-wheeled vehicles and pedestrians is often not conservative enough or lacks interpretable confidence output.

[0006] (4) Existing methods often rely on deep learning models or high-precision maps, requiring a low-computing-power, low-cost, interpretable and robust rule-driven solution, and hoping to have theoretical provability (upper bound of misjudgment, response delay, etc.). Summary of the Invention

[0007] To address the aforementioned technical problems, the present invention provides a method and system for obstacle detour determination in sanitation vehicle edge cleaning scenarios. The method, relying solely on lightweight maps (road boundaries, intersection polygons, traffic light association information) and structured perception information (obstacle position, speed, acceleration, heading, short-term predicted trajectory, taillight / traffic light status, and optional visual semantics), determines each obstacle as "bypassable" or "non-bypassable" and outputs a confidence score for downstream detour planning.

[0008] To achieve the above objectives, the technical solution of this invention provides an obstacle detour determination method for sanitation vehicle edge cleaning scenarios, comprising the following steps: S1 Feature Acquisition: Quantifying and characterizing structured obstacles, vehicle status, and lightweight map elements, outputting quantified observation sequence information including the spatial geometry and motion state of obstacles, short-term predicted trajectories, light and external signal states, and the relative relationships between obstacles and road structural elements and traffic scene nodes; S2 Intent Inference: Based on the quantified observation sequence information, generating initial intent hypotheses for each obstacle according to predicate and state transition rules, and generating corresponding rule evidence items for confirmed intents; S3 Scene Inference: Based on the quantified observation sequence information, discriminating edge cleaning-related scenes, and converting the scene discrimination results into scene intent evidence items; S4 Evidence Construction: Mapping the rule evidence items and scene intent evidence items to... S5 Evidence Fusion: Multi-source evidence is fused using evidence theory combination rules or robust combination rules, and the fused detour confidence / probability is output along with conflict and uncertainty. S6 Security Arbitration: Obstacles are judged for security based on a preset set of high-priority hard blocking rules. If any hard blocking rule is triggered, the non-detour judgment and reason are directly output to ensure that security rules are not covered by efficiency-type evidence; otherwise, proceed to step S7. S7 Temporal Stabilization: The frame-by-frame detour probability / confidence results are subjected to temporal smoothing filtering to suppress probability jitter caused by inter-frame noise, and a stable final binary judgment and confidence level are output through a hysteresis threshold mechanism. S8 Result Output: The detour judgment result is output to the outside world through the output interface module.

[0009] Further, step S2 specifically includes: S21: By adopting a finite state machine structure and driven by rule-based predicates, predefine the intention state set and its transition priority for different categories of obstacles; S22: At each moment, the state machine uses the default state corresponding to the category as the initial state and makes judgments in sequence according to the preset intention priority order; S23: For any intention state, before performing the regular trigger judgment, the state machine first judges whether the corresponding obstacle was already in the intention state at the previous moment. If the corresponding maintenance condition is met, the intention state is directly maintained and the current round of judgment ends early. If the maintenance condition is not met, the state transition judgment continues according to the trigger condition of the intention; S24: The confirmed intentions are uniformly converted into rule evidence items containing evidence strength and output.

[0010] Further, step S3 specifically includes: S31: Perform feature calculation and atomic predicate mapping, calculate the feature quantity for scene discrimination for each obstacle, and map the continuous feature quantity to a combinable atomic predicate / score; S32: Construct a group scene queue, cluster the obstacles in the corridor area in front of the vehicle in the reference line coordinate system to form a queue cluster set, and calculate the group statistics for each cluster; S33: Perform scene confidence weighted fusion, calculate the confidence of each scene according to the preset rule set and perform weighted fusion; S34: Perform scene intent mapping and output evidence, when the scene confidence meets the entry condition, output the corresponding scene intent, and use the confidence as the evidence strength.

[0011] Further, step S4 specifically includes: S41: Using {detourable, non-detourable} as the identification framework, pre-determine the pointing proposition for each piece of evidence, and configure an intrinsic strength coefficient and a confidence parameter for each piece of evidence; S42: Assign the evidence strength to the pointing proposition after being limited by the intrinsic strength coefficient, and assign the remaining quality to the uncertain terms to form a basic confidence assignment; S43: Discount and weight the basic confidence assignment through the confidence parameter to obtain a weighted basic confidence assignment; S444: Output the set of weighted basic confidence assignments.

[0012] Further, step S5 specifically includes: S51: For the evidence set, truncate and renormalize the numerical error, and retain only valid evidence with a strength greater than the threshold by setting a strength threshold; S52: Define the sum of the mass products of two pieces of evidence that support opposite propositions as the conflict degree K. When the conflict degree K of the evidence is less than the conflict threshold, normalization fusion is performed using the Dempster combination rule. When the conflict degree K of the evidence is higher than or equal to the conflict threshold, the Yager combination rule is used to recover the conflict mass to the uncertainty term; S53: Perform multi-evidence iterative folding fusion to obtain the fused mass point, and output the uncertainty and conflict degree as two auxiliary quantities; S54: Map the fused mass point to the original detour probability through BetP and output it.

[0013] Furthermore, in step S6, the preset set of high-priority hard blocking rules includes: predicted short-term intersection / minimum distance too small, vulnerable traffic participants crossing and intruding, occlusion / invisibility causing unsafe detours, conservative blocking of controlled conflicts at intersections, and insufficient geometric clearance.

[0014] Further, step S7 specifically includes: S71: after the obstacle first appears or the trajectory is interrupted for a timeout, the original detour probability is used as the initial smooth probability; S72: the smooth detour probability of the obstacle is updated recursively using the exponential weighted moving average method; S73: the final binary decision is obtained based on the hysteresis threshold and the smooth detour probability; S74: the binary decision result and confidence level are output.

[0015] Furthermore, the method also performs offline calibration of system parameters and uses an exponential sliding motion to update sensitive parameters in bounded small steps.

[0016] The technical solution of this invention also provides an obstacle detour determination system for sanitation vehicle edge cleaning scenarios, which includes the following modules: a perception and preprocessing module: quantifying and characterizing structured obstacles, vehicle status, and lightweight map elements, outputting quantified observation sequence information including the spatial geometry and motion state of obstacles, short-time predicted trajectory, light and external signal status, and the relative relationship between obstacles and road structure elements and traffic scene nodes; an initial intent state machine module: generating an initial intent hypothesis for each obstacle based on the quantified observation sequence information according to predicate and state transition rules, and generating corresponding rule evidence items for confirmed intents; a scene recognition module: judging edge cleaning-related scenes based on the quantified observation sequence information, and converting the scene judgment results into scene intent evidence items; and an evidence generation module: mapping rule evidence items to scene intent evidence items. The system assigns basic confidence and discounts and weights the basic confidence assignment based on the evidence credibility parameters obtained from offline calibration, forming an evidence set for evidence fusion. The evidence fusion module uses evidence theory combination rules or robust combination rules to fuse multi-source evidence, outputting the fused detour confidence / probability and providing conflict and uncertainty information. The conflict arbitration and safety priority module performs safety judgment on obstacles based on a preset set of high-priority hard blocking rules. If any hard blocking rule is triggered, it directly outputs the non-detour judgment and explanation, ensuring that safety rules are not covered by efficiency-related evidence. The temporal smoothing module performs temporal smoothing filtering on the frame-by-frame detour probability / confidence results to suppress probability jitter caused by inter-frame noise, and outputs a stable final binary judgment and confidence level through a hysteresis threshold mechanism. The output interface module outputs the detour judgment result to the outside world.

[0017] The technical solution of the present invention also provides a computer-readable storage medium containing a computer program, which, when executed by one or more processors, performs the obstacle detour determination method for a sanitation vehicle edge cleaning scenario as described above. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1This is a schematic diagram of the overall architecture of the obstacle detour determination system for sanitation vehicle edge cleaning scenarios of the present invention.

[0020] Figure 2 This is a flowchart illustrating the generation of rule-based evidence items in this invention;

[0021] Figure 3 This is a schematic diagram illustrating different scenarios of intent evidence output by the present invention;

[0022] Figure 4 This is a schematic diagram of the evidence generation and weighting process of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] like Figure 1As shown, the obstacle detour determination system for sanitation vehicle edge cleaning scenarios of the present invention includes the following modules: perception and preprocessing module, initial intent state machine module, scene recognition module, evidence generation module, evidence fusion module, conflict arbitration and safety priority module, timing smoothing module, and output interface module. The system works collaboratively along a chain of "feature acquisition—intent / scene inference—evidence construction—evidence fusion—security arbitration—temporal stabilization—result output": the perception and preprocessing module is responsible for quantifying and characterizing structured obstacles, vehicle states, and lightweight map elements, outputting obstacle kinematic / semantic features and their relationships with edge reference lines, intersections, and traffic lights; the initial intent state machine module forms initial intent information for obstacles based on historical window observations and predicate and state transition rules, serving as one of the sources of subsequent evidence; the scene recognition module identifies edge-clearing related scenes and calculates scene intent confidence, which participates in subsequent fusion as "scene-level evidence" without directly covering individual intents; the evidence generation module maps rule triggering degree and scene intent confidence to basic confidence allocation, and combines rule confidence parameters and observation confidence to weight evidence quality, forming evidence oriented towards the same... The system consists of a multi-source evidence set for obstacles; an evidence fusion module combines multi-source evidence based on Dempster–Shafer evidence theory or robust substitution, outputting the fused detour confidence / probability along with conflict and uncertainty; a conflict arbitration and safety priority module prioritizes hard blocking under high-risk conditions, directly outputting a non-detour decision and explanation, ensuring that safety rules are not covered by efficiency-related evidence; a temporal smoothing module performs EWMA temporal smoothing on the frame-by-frame detour probability / confidence results to reduce jitter and output a stable final binary decision and confidence level; an output interface module provides the detour probability and detour / non-detour decision for each obstacle for downstream detour planning; and a parameter calibration and adaptive mechanism runs through the above modules, calibrating and updating rule confidence, thresholds, time windows, and smoothing parameters to improve applicability under different environments and sensing conditions.

[0025] 1. Input and Quantization Definition (Feature Acquisition)

[0026] For each obstacle o, acquire and quantize at time t to form a time-series observation sequence (which may include historical windows). The quantified information includes at least: spatial geometry and motion state information of the obstacle, including position, velocity, acceleration, and heading or orientation angle; short-term prediction information, including the obstacle prediction trajectory generated within a given prediction time domain, which may be a single or multiple candidate trajectories; lighting and external signal status information, including the on / off status and temporal variation characteristics of the left turn signal, right turn signal, brake light, and hazard lights; relative relationship information with road structural elements, including the projected arc length position, lateral offset, or lateral distance of the obstacle relative to the edge reference line, and its relative front-rear position relative to the vehicle in the reference line coordinate system; and information related to traffic scene nodes, including the distance from the obstacle to the nearest intersection, stop line, or potential conflict area, and the status of the intersection traffic lights corresponding to its direction of movement; the above quantified information serves as the basic input for intent inference, evidence construction, and fusion in subsequent detour decision processing.

[0027] 2. Initial Intent State Machine

[0028] The initial intent state machine is used to generate an initial intent hypothesis for each obstacle at its current moment based on continuous time-series observations, serving as the input source for subsequent evidence construction and fusion. Driven by regularized predicates, this state machine employs a finite state machine structure, maintaining a predefined set of intent states and their transition relationships for different categories of obstacles. At each moment, the state machine uses the default state corresponding to the category as its initial state and makes judgments sequentially according to a preset intent priority order. For any intent state, before executing a regular trigger judgment, the state machine first determines whether the obstacle was already in that intent state at the previous moment. If the corresponding keep condition is met, the intent state is directly maintained, and the current round of judgment ends early, thus achieving continuous maintenance and rapid confirmation of stable intents. If the keep condition is not met, the state transition judgment continues according to the trigger condition of the intent. Through the above-mentioned "keep first, trigger second, traversal according to priority" judgment mechanism, the state machine ensures the continuity of intents while avoiding frequent state jumps and allows different intents to be extended according to a unified logic. This state machine does not output the final hard intent determination result, but generates corresponding rule evidence for the intent that is maintained or triggered, for further processing by the subsequent evidence generation and evidence fusion modules.

[0029] 2.1 Window Input and Default State

[0030] For each obstacle At any moment Set a fixed history window Quantized observation sequence Let the output state at the previous moment be... .

[0031] Initialize the current candidate state to the category default state:

[0032] Four wheels:

[0033] Two wheels:

[0034] pedestrian:

[0035] At the same time, for each non-default intent Maintenance and holding counter (Used to implement the Keep mechanism).

[0036] 2.2 General Predicate Calculation

[0037] Calculate a definite quantity within the window and use it for the triggering and holding predicates of subsequent intents:

[0038] Window average speed

[0039] Cumulative displacement of the window

[0040] Minimum horizontal distance of the window (relative to the edge reference line)

[0041] Current horizontal rate of change

[0042] Lateral offset of the predicted trajectory projected onto the reference coordinate system

[0043] : The minimum lateral distance between the predicted obstacle trajectory and the vehicle's reference line within the prediction time domain

[0044] The time index at which the minimum lateral distance is reached.

[0045] Reference line tangential heading relative to the target velocity direction .

[0046] 2.3 General "Keep-Return" State Machine Decision Process

[0047] A predefined set of intentions for the same type of obstacle. and given a fixed priority order The state machine executes intent determination sequentially according to priority at each time step, employing a decision-making mechanism combining "keep-first" and "early-return" to ensure the time consistency of intent output and suppress frequent state jitter. For any candidate intent... The determination process is as follows:

[0048] (1) Keep-first decision

[0049] If and only if the state of the previous moment At that time, perform a conditional calculation on the intention. .

[0050] Then update the hold counter. ,when At that time, it was directly confirmed as And terminate the subsequent intent determination in this round (Early-return).

[0051] like Then the counter decays. Then, the trigger determination step begins.

[0052] (2) Trigger for transition

[0053] If the decision is not made or the hold threshold is not met, calculate the trigger condition. Used for state transition decisions.

[0054] like Then set the current state to The corresponding rule evidence is written into the output set. At the same time, the "first hit locks" strategy is adopted, that is, the intent traversal is exited immediately after the hit, to avoid ambiguity caused by multiple intents hitting concurrently.

[0055] (3) Default fallback

[0056] When none of the candidate intents meet the hold threshold and no transition condition is triggered, the output category defaults to its state. .

[0057] This process uses a unified hold / trigger template for all intent states, enabling each intent to have a reusable "hold-early return-trigger transfer" mechanism, thereby achieving stable and scalable state machine decision-making in the rule-driven initial intent determination phase.

[0058] 2.4 Intent sets, priorities, and trigger / hold determination for each category

[0059] (1) Four-wheeled vehicle:

[0060] Priority order:

[0061] a. CutIn

[0062] Triggering conditions :

[0063]

[0064] A sign flip in the predicted lateral offset indicates that the predicted trajectory crosses the vehicle's reference line; or that the predicted trajectory approaches the reference line within a threshold distance in a short period of time and the arrival time does not exceed the threshold, indicating a short-term cutting trend.

[0065] Maintain conditions :

[0066]

[0067] The target will still approach the reference line quickly within the prediction domain, and the degree of approach will continue to meet a more stringent / robust threshold, which is used to maintain the cutting intention on consecutive frames.

[0068] b. Cross (crossing)

[0069] Triggering conditions :

[0070]

[0071] If the angle between the target's direction of motion and the tangential direction of the road reference line is large enough (reflecting a lateral crossing trend), and the predicted trajectory crosses the reference line, it is jointly determined to be a lateral crossing.

[0072] Maintain conditions :

[0073]

[0074] The target maintains its lateral crossing direction and continues to approach / close to the reference line corridor, avoiding frequent exits from the lateral crossing state due to short-term noise.

[0075] c. Park (parking)

[0076] Triggering conditions :

[0077]

[0078] If the target's speed and displacement remain stationary for a continuous period of time, and the duration of this stationary state exceeds the stopping threshold, it indicates that the target has entered a stable stopping state and is no longer just a short-term pause.

[0079] Maintain conditions :

[0080]

[0081] The Park state can be maintained by keeping the target in a near-static state, no longer depending on spatial position, thus avoiding frequent exits from the parking state due to slight shaking or perceived noise.

[0082] d. Temporary Stop

[0083] Triggering conditions :

[0084]

[0085] The target is currently in a near-stationary state, but the duration of stillness has not yet reached the time threshold for stable stopping, indicating that it is in a state of brief stopping or has just stopped and may be about to resume movement.

[0086] Maintain conditions :

[0087]

[0088] When the target is still in a short-term static phase and has not entered a stable stopping time range, it continues to maintain the TemporaryStop state; if the static time accumulates further and exceeds the threshold, the state machine transitions to Park.

[0089] e. Straight (default is straight)

[0090] The default straight-line state is output when no high-priority intent is triggered in the current frame.

[0091] f. Unknown

[0092] When the number of valid frames in the window is less than Output "Unknown" when a stable predicate cannot be formed.

[0093] (2) Two-wheeled vehicles:

[0094] Priority order:

[0095] a. CutIn

[0096] Trigger / hold in line with four wheels.

[0097] When the two wheels cross or rapidly approach the reference line of the vehicle, it also indicates a risk of cutting in.

[0098] b. Cross (crossing)

[0099] Trigger / hold consistent with four wheels (threshold can be set to) wait).

[0100] The two wheels made a large angle with the tangential direction of the lane and were predicted to cross the corridor, so it was determined to be a cross-traffic violation.

[0101] c. Reverse (going against the flow)

[0102] Triggering conditions :

[0103]

[0104] The direction of the speed is approximately opposite to the tangential direction of the road, and the speed is sufficient to determine that it is traveling in the wrong direction.

[0105] Maintain conditions :

[0106]

[0107] It remains in reverse motion as long as the opposite direction relationship continues and the speed is not zero.

[0108] d. Parallel (default parallel)

[0109] If a high-priority intent is not triggered, output the parallel state, indicating that the vehicle is traveling along the road direction (in the same direction or with a small deviation).

[0110] e. Unknown

[0111] Same as a four-wheeled vehicle.

[0112] (3) Pedestrians:

[0113] Priority order:

[0114] a. Cross (crossing)

[0115] Triggering conditions :

[0116]

[0117] The pedestrian's speed direction has a large angle with the road tangential direction and the lateral motion component is significant, indicating a tendency to cross the road.

[0118] Maintain conditions :

[0119]

[0120] Pedestrians should remain in a lateral crossing state while they are still within the area of ​​potential lateral movement.

[0121] b. Parallel (walking side by side along the road)

[0122] Triggering conditions :

[0123]

[0124] If a pedestrian's direction of movement is consistent with the tangential direction of the road (the deviation angle is less than the threshold) and they have a considerable walking speed, they are judged to be walking in parallel.

[0125] Maintain conditions :

[0126]

[0127] Pedestrians continue to walk in the same direction as the road and maintain a constant speed, remaining in parallel.

[0128] c. Stand (Stand still)

[0129] Triggering conditions :

[0130]

[0131] The pedestrian is almost stationary inside the window, which is determined to be standing / waiting.

[0132] Keep condition Keep(Stand)Keep(Stand)Keep(Stand):

[0133]

[0134] The Stand state is maintained as long as the static characteristic is continuously satisfied.

[0135] d. Unknown

[0136] Same as a four-wheeled vehicle.

[0137] 2.5 Rule-based Evidence Output

[0138] like Figure 2 As shown, this step does not output a hard conclusion on the intent. Instead, it converts the confirmed intents in the "initial intent state machine" (including Keep hit early return or Trigger hit and lock) into rule evidence items, which serve as the input source for subsequent evidence generation / fusion modules. For each confirmed intent... Output rule evidence items ,in For the set of propositions expressing this intent, the strength of the evidence. The continuous quantity in the triggering condition is determined by a linear saturation function. :

[0139] The strength of evidence for CutIn is determined by both "lateral proximity" and "time decay after triggering":

[0140]

[0141] in This indicates the lateral nearest distance estimation. The CutIn distance threshold; The time interval from the trigger time. For time scale, This is the attenuation coefficient.

[0142] The strength of evidence for the cross is obtained by continuously mapping the yaw difference between the velocity direction and the reference direction:

[0143]

[0144] in The velocity direction angle, Reference direction angle; The initial threshold for Cross is... This is the maximum normalization threshold.

[0145] Park / Stop: Use The linear saturation mapping combination is obtained (corresponding to the threshold in the triggering condition).

[0146] Output This step serves as a source of evidence for subsequent evidence generation / fusion; it does not output a hard conclusion for the system's final detour decision.

[0147] 3. Scene recognition (soft entry)

[0148] In the context of edge cleaning, traffic / operation scenarios are identified based on temporal quantitative observations of obstacles, short-term prediction information, and lightweight map elements. The scenario identification results are then converted into scenario intent evidence that can be used for evidence fusion. Let the scenario set be... and define the scene intent set. Wherein: Pullover represents the context of "parking close to the side of the road"; Queue represents the context of "queue constraints / waiting in line"; and Block represents the context of "occlusion by large objects or restricted passage". The scene recognition output adopts a soft entry mechanism, and the scene intent participates in subsequent fusion as evidence.

[0149] 3.1 Feature Calculation and Atomic Predicates

[0150] For each obstacle At any moment Calculate the features used for scene discrimination (current frame rate and / or window statistics are both acceptable), including at least: average velocity. Cumulative displacement acceleration ; Lateral distance relative to the edge reference line and the minimum horizontal distance of the prediction domain Distance to intersection / stop line / potential conflict zone and whether it is in an intersection area Lighting status Category and size ; and the state associated with the phase of the traffic lights. .

[0151] Mapping continuous quantities to composable atomic predicates / scores (example):

[0152]

[0153] in For linear saturation to The mapping; the threshold and weight are calibrable parameters.

[0154] 3.2 Queue Construction in Group Scenarios

[0155] For scenarios requiring multiple obstacles to be formed together, a candidate queue is first constructed. The projected arc length of each obstacle is then calculated in the reference line coordinate system. With lateral offset Select the area in front of the vehicle's aisle: ,Will Internal target Sort by vertical interval threshold Segmented clustering yields a set of queue clusters. Calculate population statistics for each cluster:

[0156] ,when Not less than the minimum quantity threshold When conditions are met, the formation of a Queue-like scenario is allowed; otherwise, Slow-traffic and Junction-waiting are not determined.

[0157] 3.3 Scene Rule Set and Confidence Calculation

[0158] Configure a set of rules for each type of scenario Each rule outputs The rule score is obtained by weighted fusion of confidence scores for each scenario.

[0159]

[0160] in" "Indicates the object of the rule: Pullover-scene and Large-vehicle-scene, with a single obstacle." For objects; Slow-traffic-scene and Junction-waiting-scene are clustered in queues. For objects. Example rules are as follows:

[0161] Pullover-scene (object: It is characterized by the degree of proximity to the edge, low-speed characteristics, edge-fitting prediction, and light signals.

[0162]

[0163] Slow-traffic-scene (object: Characterized by queue size, slow speed, and queue length / density.

[0164]

[0165] Junction-waiting-scene(object: It is characterized by intersection proximity, the proportion of stationary elements within a cluster, and signal control features.

[0166]

[0167] in The percentage of low-speed / stationary targets within the cluster. This represents the distance between the cluster and the stop line or conflict area.

[0168] Large-vehicle-scene (object: (Can be independent of multiple targets): Characterized by the attributes of large targets, relative corridor position, and relative distance.

[0169]

[0170] 3.4 Mapping from Scene Confidence to Scene Intent and Evidence Output

[0171] Set scene entry threshold .like Figure 3 As shown, when the scene confidence level meets the entry condition, the corresponding scene intent is output, and the confidence level is used as the strength of evidence. The mapping relationship is as follows:

[0172]

[0173] If the same obstacle simultaneously meets the entry conditions of multiple scenarios, multiple scenario intent evidence items are allowed to be output; alternatively, the maximum confidence priority strategy can be adopted to output only the scenario intent evidence item with the strongest strength.

[0174] Finally, for each obstacle Output the set of evidence for the scene's intent: And use it as input for subsequent evidence construction and evidence fusion.

[0175] 4. Evidence Generation and Weighting

[0176] For the same obstacle at any time All types of evidence obtained (including rule-based evidence and scenario-based evidence) are uniformly converted into an evidence theory identification framework. The basic confidence assignment is then applied, and the basic confidence assignment is discounted and weighted based on the evidence credibility parameters obtained from offline calibration, forming an evidence set that can be directly used for evidence fusion.

[0177] 4.1 Evidence Items and Identification Framework

[0178] For each obstacle At any moment The evidence items are represented as follows:

[0179]

[0180] in This serves as the intent / scenario intent identifier corresponding to the evidence item. This represents the strength of the evidence. Define the identification framework:

[0181]

[0182] And allows the assignment of uncertainty quality to .

[0183] For each Presuppose its pointing proposition (polarity):

[0184]

[0185] For each Configure intrinsic strength coefficient With credibility parameters The parameters can be obtained through offline calibration.

[0186] 4.2 Construction of Basic Confidence Assignment

[0187] For evidence items Constructing the basic confidence assignment :

[0188]

[0189] That is: the strength of evidence is determined by After being limited, the mass is allocated to the proposition it points to, and the remaining mass is allocated to the uncertain term. .

[0190] 4.3 Credibility Discount Weighting

[0191] credibility parameter Applying the basic confidence assignment to the base results in the discounted weighted basic confidence assignment. :

[0192]

[0193] 4.4 Output

[0194] For each obstacle Output the weighted set of basic confidence assignments:

[0195]

[0196] The set As input to the subsequent evidence fusion unit. Figure 4 As shown.

[0197] 5. Evidence Fusion and Conflict Resolution

[0198] For the same obstacle At any moment The evidence set is combined using the evidence theory combination rules to combine the multi-source basic confidence assignments, resulting in a fusion mass point. Explicitly calculate conflict degree during the fusion process. With uncertainty When the conflict level exceeds a threshold, the system switches to a robust combination rule (such as the Yager rule) to avoid misjudgments caused by over-normalization, and maps the fusion result to the original detour probability using Belief / Plausibility or BetP. .

[0199] 5.1 Input Integration and Standardization

[0200] Set up an identification framework For obstacles At any moment The set of evidence is denoted as

[0201] Each of them For the basic confidence assignment, satisfying

[0202]

[0203] Numerical errors can be truncated and renormalized (for example, negative values ​​can be truncated to 0 and then normalized to the sum).

[0204] To reduce the perturbation of fusion by invalid evidence, a strength threshold can be set. Only retain those that meet the requirements. Evidence is involved in the fusion.

[0205] 5.2 Two-Evidence Combination and Conflict Calculation under the Binary Framework

[0206] For any two pieces of evidence Its degree of conflict is defined as the sum of the products of the masses of mutually supporting opposite propositions:

[0207]

[0208] (1) Dempster combination (low conflict case)

[0209] when ( When the conflict threshold is reached, Dempster normalized combination is used to obtain... :

[0210]

[0211] Within the binary framework, this can be written in a directly implementable closed form:

[0212]

[0213] in It means Bypassable. This indicates Non-Bypassable.

[0214] (2) Yager combination (high conflict robust case)

[0215] when At that time, the Yager rule is used to recover conflicting mass into the uncertain term. To avoid "strong normalization" amplifying conclusions from one side:

[0216]

[0217] The same direct implementation of the binary framework can also be used:

[0218]

[0219] 5.3 Iterative Fusion of Multiple Evidences and Output of Global Conflict / Uncertainty

[0220] right In Iterative folding and fusion of evidence:

[0221]

[0222] The final fused particle is obtained:

[0223]

[0224] Simultaneously output two types of auxiliary quantities for subsequent decision-making:

[0225] Uncertainty: ;

[0226] Conflict degree: can be the maximum conflict or the cumulative conflict during the iteration process (either can be chosen and fixed in the implementation), for example

[0227]

[0228] High conflict This usually increases significantly, thus naturally manifesting as "prudent decision-making".

[0229] 5.4 Mapping from fused particles to Belief / Plasusibility and detour probability

[0230] Within the binary framework, Belief and Plausibility can be directly derived from the point mass:

[0231]

[0232] Map the fusion result to the original detour probability. A feasible and commonly used approach is BetP (pignistic transform):

[0233]

[0234] Alternatively, a range-based output can be used for conservative decision-making: As the "circumvention probability interval", the downstream strategy selects the more conservative side.

[0235] 5.5 Output

[0236] Output fused mass and derived quantity:

[0237]

[0238] in (Or output both types of probabilities simultaneously) for subsequent use by the security arbitration and decision-making module.

[0239] 6. Safety Priority and Hard Blocking Rules

[0240] For each obstacle At any moment A set of high-priority hard blocking rules is preset. The rules are triggered solely based on structured kinematic information, short-term predicted trajectories, reference lines / intersection map elements, and visibility geometric criteria. When any hard blocking rule meets the triggering condition, an unavoidable detour determination is directly output, and the reason code and key trigger quantity are reported, thereby ensuring that safety-related conclusions are not overridden by other evidence.

[0241] Hard blocking rules include, but are not limited to:

[0242] (1) Predicting short-term intersection / minimum distance is too small: If the minimum predicted distance between the vehicle and the target in the prediction time domain is too small. And the corresponding time (or If ), then a hard block is triggered;

[0243] (2) Vulnerable road users crossing and intruding: If Furthermore, if the predicted trajectory crosses the reference line corridor (e.g., a lateral sign flip occurs) and approaches the corridor to the threshold within a short time window, a hard block is triggered.

[0244] (3) Obstruction / invisibility makes safe detour impossible: If the proposed detour corridor or conflict area is not observable ( If the target is within close-range influence, a hard block is triggered;

[0245] (4) Controlled conflict at intersection: If the target is in the vicinity of the intersection / stop line and the traffic light is red or unknown, a hard block is triggered.

[0246] (5) Insufficient geometric clearance: If the remaining clearance for detour is calculated based on the road boundary and the target geometry. This triggers a hard block.

[0247] When any make At that time, output directly

[0248]

[0249] Output the reason information. (For example (This is for downstream planning and interpretation purposes.)

[0250] 7. Timing smoothing and final decision

[0251] For each obstacle At any moment Original detour probability Temporal filtering is performed to suppress probability jitter caused by inter-frame noise and stabilize the final decision. A smooth probability is maintained for each obstacle. Compared with the previous moment .

[0252] When the obstacle first appears or the trajectory is interrupted for more than [time period]... Initialization:

[0253]

[0254] Subsequently, an exponentially weighted moving average (EWMA) is used for recursive updates:

[0255]

[0256] To avoid frequent flipping around the threshold, a hysteresis threshold is introduced. And output the binary decision according to the following rules:

[0257]

[0258] in It means Bypassable. This indicates Non-Bypassable.

[0259] Final output judgment and confidence level:

[0260]

[0261] And provide it to downstream detour planning calls.

[0262] 8. Parameter Calibration and Adaptation

[0263] To ensure the stability and effectiveness of the rules and judgment logic under different road and operational environments, the system parameters are calibrated offline and updated adaptively in small steps online. The set of parameters to be calibrated includes: the credibility of each piece of evidence. Evidence strength coefficient Various threshold parameters (distance / time / speed thresholds, etc.), historical window length Prediction time domain and timing smoothing parameters With hysteresis threshold .

[0264] Offline calibration based on historical labeled datasets Perform: For each type of evidence The proportion of samples that match the label on the trigger samples is obtained by statistical analysis. (Or use Bayesian estimation with beta prior to improve robustness for small samples); use grid / random search for threshold and window parameters, and select the optimal parameter set under the objectives of "false positive rate (paying a higher cost for errors that classify non-bypassable as bypassable) + output stability (number of flips)"; determine recommended values ​​for smoothing and hysteresis parameters based on the criterion of "minimum jitter and not too slow response".

[0265] Online adaptive updates perform bounded small-step updates on a small number of sensitive parameters using an exponential sliding motion approach.

[0266]

[0267] in The target value is obtained from the statistics of the most recent time window. For learning rate, To establish safety upper and lower bounds, parameters are used to prevent drift and out-of-bounds errors. The calibrated parameters are then used to update rule confidence, thresholds, windows, and smoothing parameters, thereby improving portability and consistency across different environments.

[0268] The method of this invention, relying solely on lightweight maps (road boundaries, intersection polygons, traffic light association information) and structured perception information (obstacle position, speed, acceleration, heading, short-term predicted trajectory, taillight / traffic light status, and optional visual semantics), determines each obstacle as "bypassable" or "non-bypassable" and outputs a confidence score for downstream detour planning. This invention provides a low-computing-power, low-cost, interpretable, and robust rule-driven solution with theoretical provability (upper bound on misjudgment, response delay, etc.).

[0269] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for determining obstacle detour in a sanitation vehicle edge-sweeping scenario, characterized in that, Includes the following steps: S1 Feature Acquisition: Quantify and characterize structured obstacles, vehicle status, and lightweight map elements, and output quantitative observation sequence information including the spatial geometry and motion state of obstacles, short-term predicted trajectory, light and external signal status, and the relative relationship between obstacles and road structure elements and traffic scene nodes. S2 Intent Inference: Based on quantified observation sequence information, an initial intent hypothesis is generated for each obstacle according to predicate and state transition rules, and corresponding rule evidence items are generated for the confirmed intent; S3 Scene Inference: Based on quantitative observation sequence information, the scene related to edge cleaning is identified, and the scene identification results are converted into scene intent evidence items; S4 Evidence Construction: Map rule-based evidence items and scene intent evidence items to basic confidence assignments, and apply discount weights to the basic confidence assignments based on the evidence credibility parameters obtained from offline calibration to form an evidence set for evidence fusion; S5 Evidence Fusion: It uses evidence theory combination rules or robust combination rules to fuse multi-source evidence, outputs the fused detour confidence / probability and gives the conflict and uncertainty; S6 Safety Arbitration: Based on the preset set of high-priority hard blocking rules, the safety of the obstacle is determined. If any hard blocking rule is triggered, the determination that the obstacle cannot be bypassed and the reason are directly output to ensure that the safety rules are not covered by efficiency-related evidence. Otherwise, proceed to step S7. S7 temporal stability: The frame-by-frame bypass probability / confidence results are subjected to temporal smoothing filtering to suppress probability jitter caused by inter-frame noise, and a stable final binary judgment and confidence level are output through the hysteresis threshold mechanism. S8 Result Output: Output the bypass judgment result to the outside world through the output interface module.

2. The method according to claim 1, characterized in that, Step S2 specifically includes: S21: By adopting a finite state machine structure and using regularized predicates as the driving force, predefine the intention state set and its transition priority for different types of obstacles; S22: At each moment, the state machine uses the default state corresponding to the category as the initial state and makes judgments in sequence according to the preset intention priority order; S23: For any intention state, before executing the regular trigger judgment, the state machine first judges whether the corresponding obstacle was already in the intention state in the previous moment. If the corresponding holding condition is met, the intention state is directly held and the current round of judgment ends early. If the holding condition is not met, the state transition judgment continues according to the trigger condition of the intention. S24: Convert the confirmed intents into rule-based evidence items that include the strength of evidence and output them.

3. The method according to claim 1, characterized in that, Step S3 specifically includes: S31: Perform feature calculation and atomic predicate mapping, calculate the feature quantity for scene discrimination for each obstacle, and map the continuous feature quantity to a combinable atomic predicate / score; S32: Construct a group scene queue, cluster obstacles in the corridor area in front of the vehicle in the reference line coordinate system to form a queue cluster set, and calculate the group statistics for each cluster; S33: Perform scene confidence weighted fusion, calculate the confidence of each scene according to the preset rule set and perform weighted fusion; S34: Perform scene intent mapping and output evidence. When the scene confidence meets the entry condition, output the corresponding scene intent and use the confidence as the evidence strength.

4. The method according to claim 1, characterized in that, Step S4 specifically includes: S41: Using {detourable, non-detourable} as the identification framework, a pre-defined proposition is set for each piece of evidence, and an intrinsic strength coefficient and credibility parameter are configured for each piece of evidence; S42: The strength of evidence is limited by the intrinsic strength coefficient and then allocated to the directional proposition. The remaining mass is allocated to the uncertain terms to form a basic confidence allocation. S43: Discount and weight the basic confidence assignment using the confidence parameter to obtain the weighted basic confidence assignment; S43: Output the weighted set of basic confidence assignments.

5. The method according to claim 1, characterized in that, Step S5 specifically includes: S51: For the evidence set, the numerical error is truncated and renormalized, and by setting an intensity threshold, only valid evidence with an intensity greater than the threshold is retained. S52: The sum of the products of the qualities of two pieces of evidence that support opposite propositions is defined as the degree of conflict K. When the degree of conflict K of the evidence is less than the conflict threshold, the Dempster combination rule is used for normalization fusion. When the degree of conflict K of the evidence is higher than or equal to the conflict threshold, the Yager combination rule is used to recover the conflict quality to the uncertain term. S53: Perform multi-evidence iterative folding and fusion to obtain the fused mass point, and output uncertainty and conflict degree as two auxiliary quantities; S54: Map the fused particles to the original detour probabilities using BetP and output them.

6. The method according to claim 1, characterized in that, In step S6, the preset set of high-priority hard blocking rules includes: predicted short-term intersection / minimum distance too small, vulnerable traffic participants crossing and intruding, occlusion / invisibility causing unsafe detours, conservative blocking of controlled conflicts at intersections, and insufficient geometric clearance.

7. The method according to claim 1, characterized in that, Step S7 specifically includes: S71: After the obstacle first appears or the trajectory is interrupted for a timeout, the original detour probability is used as the initial smoothing probability; S72: The smooth bypass probability of obstacles is updated recursively using the exponentially weighted moving average method; S73: The final binary decision is obtained based on the hysteresis threshold and the smooth detour probability; S74: Output the binary decision result and confidence level.

8. The method according to claim 1, characterized in that, The method also performs offline calibration of system parameters and uses an exponential sliding motion to update sensitive parameters in bounded small steps.

9. An obstacle avoidance determination system for sanitation vehicle edge cleaning scenarios, characterized in that, Includes the following modules: Perception and preprocessing module: Quantifies and characterizes structured obstacles, vehicle status and lightweight map elements, and outputs quantitative observation sequence information including the spatial geometry and motion state of obstacles, short-term predicted trajectory, light and external signal status, and the relative relationship between obstacles and road structure elements and traffic scene nodes. Initial Intent State Machine Module: Based on quantized observation sequence information, it generates an initial intent hypothesis for each obstacle according to predicate and state transition rules, and generates corresponding rule evidence items for the confirmed intent; Scene recognition module: Based on quantified observation sequence information, it identifies scenes related to edge cleaning and converts the scene identification results into scene intent evidence items; Evidence generation module: Maps rule-based evidence items and scene intent evidence items to basic confidence assignments, and applies discount weights to the basic confidence assignments based on the evidence credibility parameters obtained from offline calibration to form an evidence set for evidence fusion; Evidence Fusion Module: It uses evidence theory combination rules or robust combination rules to fuse multi-source evidence, outputs the fused bypass confidence / probability, and provides conflict and uncertainty; Conflict Arbitration and Safety Priority Module: Based on a preset set of high-priority hard blocking rules, the module makes safety judgments on obstacles. If any hard blocking rule is triggered, it directly outputs the determination that the obstacle cannot be bypassed and the reason, so as to ensure that the safety rules are not covered by efficiency-related evidence. Temporal smoothing module: Performs temporal smoothing filtering on the frame-by-frame bypass probability / confidence results to suppress probability jitter caused by inter-frame noise, and outputs a stable final binary judgment and confidence level through a hysteresis threshold mechanism; Output interface module: The output interface module outputs the bypass judgment results to the outside world.

10. A computer-readable storage medium containing a computer program, characterized in that, When the computer program is executed by one or more processors, it performs the obstacle detour determination method for the sanitation vehicle edge cleaning scenario as described in any one of claims 1-8.

Images