CN-122009162-A - Obstacle detour judging method and system for sanitation vehicle edge cleaning scene

CN122009162ACN 122009162 ACN122009162 ACN 122009162ACN-122009162-A

Abstract

The invention discloses an obstacle detouring judging method and system for a sanitation truck edge cleaning scene, wherein the method is used for judging 'detourable (Bypassable)' or 'Non-detourable (Non-Bypassable)' for each obstacle under the condition of only relying on a lightweight map (road boundary, intersection polygon and signal lamp related information) and structural perception information (obstacle position, speed, acceleration, heading, short-term prediction track, tail lamp/signal lamp state and optional visual semantics), and outputting confidence level for downstream detouring planning.

Inventors

HE TAO
LIAO WENLONG
ZHOU HUI
ZHANG RUNXI

Assignees

酷哇科技有限公司
上海酷移机器人有限公司

Dates

Publication Date: 20260512
Application Date: 20260211

Claims (10)

1. The obstacle detour judging method for the sanitation truck edge cleaning scene is characterized by comprising the following steps of: S1, obtaining characteristics, namely quantifying and characterizing structural barriers, vehicle states and lightweight map elements, and outputting quantified observation sequence information comprising space geometry and motion states of the barriers, short-time predicted tracks, light and external signal states and relative relations between the barriers, road structural elements and traffic scene nodes; S2, intention inference, namely generating an initial intention hypothesis for each obstacle according to predicate and state transition rules based on quantized observation sequence information, and generating a corresponding rule evidence item for the confirmed intention; S3, scene estimation, namely judging related scenes along the edge cleaning based on quantized observation sequence information, and converting scene judging results into scene intention evidence items; S4, evidence construction, namely mapping rule evidence items and scene intention evidence items into basic confidence allocation, and discount weighting the basic confidence allocation according to evidence credibility parameters obtained through offline calibration to form an evidence set for evidence fusion; s5, evidence fusion, namely fusing the multi-source evidence by adopting an evidence theory combination rule or a robust combination rule, outputting the bypass confidence/probability after fusion, and giving out conflict and uncertainty; S6, safety arbitration, namely carrying out safety judgment on the obstacle based on a preset high-priority hard blocking rule set, if any hard blocking rule is triggered, directly outputting an unreliability bypass judgment and a reason explanation to ensure that the safety rule is not covered by efficiency type evidence, otherwise, continuing to step S7; S7, time sequence stabilization, namely performing time sequence smoothing filtering on the frame-by-frame bypass probability/confidence result to inhibit probability jitter caused by inter-frame noise, and outputting stable final binary judgment and confidence through a hysteresis threshold mechanism; And S8, outputting a result, namely outputting a detourable judgment result through an output interface module.
2. The method according to claim 1, wherein step S2 specifically comprises: s21, respectively predefining an intention state set and a transfer priority thereof for different types of obstacles by adopting a finite state machine structure and taking a regularized predicate as a drive; s22, the state machine takes a default state corresponding to the category as an initial state at each moment, and sequentially judges according to a preset intention priority order; S23, for any intention state, the state machine firstly judges whether the corresponding obstacle is in the intention state at the last moment before executing the conventional trigger judgment, if the corresponding holding condition is met, the intention state is directly held and the current round of judgment is finished in advance, and if the holding condition is not met, the state transition judgment is continuously carried out according to the intention trigger condition; And S24, uniformly converting the confirmed intents into regular evidence items containing evidence intensity and outputting the regular evidence items.
3. The method according to claim 1, wherein step S3 specifically comprises: S31, performing feature calculation and atomic predicate mapping, calculating feature quantities for scene discrimination for each obstacle, and mapping continuous feature quantities into combinable atomic predicates/scores; S32, constructing a group scene queue, clustering obstacles in the area of a front corridor of a host computer under a reference line coordinate system to form a queue cluster set, and calculating group statistics for each cluster; s33, carrying out scene confidence degree weighted fusion, calculating the confidence degree of each scene according to a preset rule set, and carrying out weighted fusion; And S34, mapping scene intention and outputting evidence, and outputting corresponding scene intention when the field Jing Zhixin degrees meet the entering condition, wherein the confidence is used as the evidence intensity.
4. The method according to claim 1, wherein step S4 specifically comprises: S41, presetting a pointing proposition for each evidence item by taking { detourable, unreliable } as an identification frame, and configuring an intrinsic strength coefficient and a credibility parameter for each evidence item; S42, distributing evidence strength to a pointing proposition after limiting by an intrinsic strength coefficient, and distributing the residual quality to an uncertain item to form basic confidence distribution; S43, discount weighting is carried out on the basic confidence allocation through the credibility parameter, and the weighted basic confidence allocation is obtained; S43, outputting the weighted basic confidence allocation set.
5. The method according to claim 1, wherein step S5 specifically comprises: s51, cutting off and renormalizing the numerical error aiming at the evidence set, and only reserving effective evidence with the intensity larger than a threshold by setting the intensity threshold; S52, defining the sum of quality products of two evidences supporting opposite propositions as conflict degree K, adopting a Dempster combination rule to perform normalized fusion when the conflict degree K of the evidences is smaller than a conflict threshold, and adopting Yager combination rules to recover conflict quality to an uncertain item when the conflict degree K of the evidences is higher than or equal to the conflict threshold; S53, performing multi-evidence iterative folding fusion to obtain fusion particles, and outputting uncertainty and conflict degree as two auxiliary quantities; S54, the fusion particles are mapped into the original detour probability through BetP and output.
6. The method of claim 1, wherein the set of preset high priority hard-stop rules in step S6 includes predicting that short-term intersections/minimum distances are too small, that disadvantaged traffic participants cross intrusions, that occlusion/invisibility results in unsafe detours, that intersection controlled conflict conservative stops, and that geometric gaps are insufficient.
7. The method according to claim 1, wherein step S7 specifically comprises: S71, taking the original detour probability as the initial smoothing probability after the first occurrence of the obstacle or the overtime of the track interruption; S72, recursively updating smooth detour probability of the obstacle by adopting an exponential weighted moving average method; S73, acquiring final binary judgment based on a hysteresis threshold and smooth detour probability; S74, outputting the binary judgment result and the confidence coefficient.
8. The method of claim 1, further comprising calibrating system parameters offline and performing bounded small-step updates of sensitive parameters using an exponential sliding form.
9. The obstacle detouring judging system for the sanitation truck edge cleaning scene is characterized by comprising the following modules: The sensing and preprocessing module is used for quantifying and characterizing structural barriers, vehicle states and lightweight map elements and outputting quantitative observation sequence information comprising space geometry and motion states of the barriers, short-time prediction tracks, light and external signal states and relative relations between the barriers, road structural elements and traffic scene nodes; an initial intention state machine module, which generates an initial intention hypothesis for each obstacle according to predicate and state transition rules based on the quantized observation sequence information, and generates a corresponding rule evidence item for the confirmed intention; The scene recognition module is used for judging the related scene along the edge cleaning based on the quantized observation sequence information and converting a scene judgment result into a scene intention evidence item; The evidence generation module is used for mapping the rule evidence items and the scene intention evidence items into basic confidence allocation, and discount weighting is carried out on the basic confidence allocation according to the evidence credibility parameters obtained through offline calibration to form an evidence set for evidence fusion; The evidence fusion device module is used for fusing the multi-source evidence by adopting an evidence theory combination rule or a robust combination rule, outputting the fused bypass confidence/probability and giving out conflict and uncertainty; The conflict arbitration and safety priority module is used for carrying out safety judgment on the obstacle based on a preset high-priority hard blocking rule set, and if any hard blocking rule is triggered, directly outputting the unreliability bypass judgment and the reason explanation so as to ensure that the safety rule is not covered by the efficiency evidence; The time sequence smoothing module is used for performing time sequence smoothing filtering on the frame-by-frame bypass probability/confidence result to inhibit probability jitter caused by inter-frame noise, and outputting stable final binary judgment and confidence through a hysteresis threshold mechanism; and the output interface module is used for outputting the detourable judging result outwards.
10. A computer-readable storage medium containing a computer program, which, when executed by one or more processors, performs the obstacle detour determination method of the sanitation-oriented truck-edge cleaning scenario of any one of claims 1-8.

Description

Obstacle detour judging method and system for sanitation vehicle edge cleaning scene Technical Field The invention relates to the technical field of automatic driving, in particular to an obstacle detouring judging method and system for a sanitation truck edge cleaning scene. Background When the sanitation truck performs the edge cleaning on the urban road, the sanitation truck is often required to run close to the right side edge of the road. The road edge frequently presents various dynamic/static obstacles such as side-by-side parking, approaching goods taking, short stopping, waiting in line for vehicles, two-wheelers, pedestrians and the like. The existing methods generally suffer from the following disadvantages: (1) Most are based on simple hard rules (if-else) or single threshold coverage, lack of modeling of observation uncertainty, and are prone to false positives under perceived noise. (2) Scene recognition often employs hard overlay policies (scene is immediately overlaid with individual intent), resulting in false overlays and efficiency loss. (3) Treatment of two-wheelers and pedestrians is often not conservative enough or lacks an interpretable confidence output. (4) Existing methods often rely on deep learning models or high precision maps, require a low-computational-effort, low-cost, interpretable and robust rule-driven scheme, and hope to have theoretical provability (upper bound of erroneous judgment, response delay, etc.). Disclosure of Invention In order to solve the technical problems, the technical scheme of the invention provides an obstacle detouring judging method and system for a sanitation truck edge cleaning scene, wherein the method judges 'detourable (Bypassable)' or 'Non-detourable (Non-Bypassable)' for each obstacle under the condition of only relying on a lightweight map (road boundary, intersection polygon, signal lamp associated information) and structured perception information (obstacle position, speed, acceleration, heading, short-term prediction track, tail lamp/signal lamp state and optional visual semantics), and outputs a confidence level for downstream detouring planning. In order to achieve the aim, the technical scheme of the invention provides an obstacle detouring judging method for a sanitation truck edge cleaning scene, which comprises the following steps of S1 feature acquisition, S4 evidence construction, wherein the S1 feature acquisition comprises quantification and characterization of structural obstacles, truck states and lightweight map elements, the output comprises quantification observation sequence information of spatial geometry and motion states of the obstacles, short-time prediction tracks, light and external signal states and relative relations between the obstacles and road structure elements and traffic scene nodes, S2 intention inference comprises the steps of generating initial intention hypothesis for each obstacle according to predicate and state transfer rules and generating corresponding rule evidence items for the confirmed intention according to predicate and state transfer rules, S3 scene inference comprises the steps of judging related edge cleaning scenes based on quantification observation sequence information, converting scene judgment results into scene intention evidence items, S4 evidence construction comprises the steps of mapping the rule evidence items and the scene intention evidence items into basic confidence allocation, weighting the basic confidence allocation according to offline obtained evidence reliability parameters, forming a evidence fusion set, S5 comprises the steps of adopting a combination rule or adopting a combination rule to judge the basic confidence allocation, and giving out a high-level evidence fusion rule, and a high-level evidence detouring effect is not guaranteeing the safety rule-round if the safety rule is not triggered by the combination rule or the combination rule is arranged on the basis of the high-priority safety rule and the safety rule is not triggered, otherwise, the step S7 is continued, the step S7 is carried out in a time sequence stabilizing mode, time sequence smoothing filtering is carried out on the frame-by-frame bypass probability/confidence result to restrain probability jitter caused by inter-frame noise, stable final binary judgment and confidence degree are output through a hysteresis threshold mechanism, and the step S8 is carried out in a result output mode, namely a bypass judgment result is output through an output interface module. Further, the step S2 specifically comprises the steps of S21 of respectively predefining an intention state set and a transition priority thereof for different types of barriers by adopting a finite state machine structure and taking a regularized predicate as a drive, S22 of sequentially judging by a state machine at each moment by taking a default state corresponding to the type as an initial state according to a preset intention prior