Three-subject examination safety protection system and method based on vision and radar fusion

Abstract

The invention discloses a third subject examination safety protection system and method based on vision and radar fusion, and the method comprises the following steps: obtaining millimeter wave radar, vision, ultrasonic and laser radar data, completing the alignment verification, and forming collaborative perception data and availability states; constructing an examination area, a dangerous road section and time section risk priori, and generating a scene constraint; fusing the collaborative sensing data, the availability state and the scene constraint, and outputting fusion evidence; recurring the target set under the prior constraint, identifying pedestrians and obstacles, and generating spatial relationship representation; evaluating a pedestrian priority danger level and controlling braking; and synchronously retaining the risk event and adaptively updating the configuration according to the evidence receipt to form a safety protection closed loop. Based on multi-sensor evidence fusion and a scene prior constraint method, the pedestrian priority safety protection closed loop in the third subject examination is achieved, and the method has the advantages of being stable in recognition, reliable in braking and high in self-adaptive correction capacity.

Description

Technical Field

[0001] This invention relates to the field of active safety control for vehicles, and in particular to a safety protection system and method for driving test subject 3 based on vision and radar fusion. Background Technology

[0002] Existing driver safety assistance and training vehicle protection systems typically employ multi-source sensors such as millimeter-wave radar and cameras to detect and fuse road targets. Some solutions incorporate depth vision or lidar to improve target recognition accuracy and trigger deceleration or braking control when potential collision risks are detected. In the driving test scenario, the system may also combine test area information, road markings, and test route to impose scenario constraints, and record key risk events for post-event review and management, thereby achieving safe intervention and evidence archiving during the examination process.

[0003] However, existing technologies often lack evidence discounting and fusion adjudication mechanisms that are tied to sensor availability when there are inconsistencies in the timing of multi-source sensing, signal anomalies, and strong evidence conflicts. This makes it difficult for fusion confidence and conflict levels to stably constrain subsequent target generation and tracking, leading to false detections, missed detections, or short-term jitter triggering braking. Furthermore, risk event retention is mostly limited to the recording level, lacking the ability to generate evidence receipts from background processing actions and drive configuration updates in a closed loop. This makes it difficult to continuously correct fusion strategies and gating thresholds, resulting in insufficient adaptability of the system to changes in risk levels in different areas, dangerous road sections, and time periods.

[0004] Therefore, how to provide a safety protection system and method for the driving test based on vision and radar fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a safety protection system and method for the driving test (subject 3) based on vision and radar fusion. This invention is based on multi-sensor evidence fusion and scene prior constraint method to realize a closed loop of pedestrian priority safety protection for the driving test (subject 3), and has the advantages of stable recognition, reliable braking and strong adaptive correction capability.

[0006] The safety protection method for driving test subject 3 based on vision and radar fusion according to an embodiment of the present invention includes the following steps:

[0007] Acquire environmental perception data and perform time alignment and signal verification to form a multi-sensor collaborative perception data set, and generate a sensor availability status set;

[0008] Based on the multi-sensor collaborative perception data set, a prior set of examination area, a prior set of dangerous road section, and a prior set of time period risk are constructed, and then fused to generate a prior constraint set of examination scenario;

[0009] The multi-sensor collaborative sensing data set, the sensor availability status set, and the examination scenario prior constraint set are input into the signal analysis host, which performs evidence theory fusion adjudication and outputs a fusion evidence result set.

[0010] Based on the fusion evidence result set, the target generation gating is performed to obtain evidence through the target set. Under the constraints of the prior constraint set of the examination scenario, the evidence credibility constraint random finite set target set is recursively calculated. Pedestrian target recognition results, obstacle target recognition results and target spatial relationship representation are extracted to form a fusion environment representation set.

[0011] Based on the fusion of environmental characterization sets, pedestrian priority hazard level analysis is performed, hazard level results and trigger evidence retention instructions are output and sent to the braking control host to implement deceleration braking or emergency braking, forming braking execution results;

[0012] The system sends the trigger evidence retention instruction to the vehicle monitoring and transmission host to generate risk event retention results. Based on the risk event retention results and the background processing actions, it performs adaptive discount updates to complete the closed loop of exam security protection.

[0013] Optionally, the generation of the multi-sensor collaborative sensing data set and the sensor availability status set specifically includes:

[0014] Simultaneously acquire millimeter-wave radar data, depth vision sensor data, high-definition camera data, ultrasonic radar data, and lidar data at the test vehicle end to form a raw set of environmental perception data;

[0015] Perform time alignment processing on the raw set of environmental perception data, and output a time-aligned data set;

[0016] Perform signal verification processing on the time-aligned data set and output a set of signal verification results.

[0017] Based on the signal verification result set, a corresponding sensor availability status record is generated for each sensor in each alignment window, and a sensor availability status set is constructed, which includes online status items, data integrity items, measurement stability items and consistency conclusion items.

[0018] Based on the sensor availability status set, the time-aligned data set is subjected to availability filtering and back-off encapsulation to obtain a multi-sensor collaborative sensing data set.

[0019] Optionally, the generation of the prior constraint set for the examination scenario specifically includes:

[0020] The location and driving status information of the test vehicle are obtained based on the multi-sensor collaborative perception data set, and spatial matching is performed with the pre-configured test area boundary data to generate a set of test vehicle area affiliation results.

[0021] Based on the test vehicle area attribution result set, the multi-sensor collaborative perception data set is grouped by region to generate a region-grouped collaborative perception data set;

[0022] Using the region identifier as an index, count the historical risk events, historical braking execution counts, and historical pedestrian priority trigger counts within the region, and output the set of regional statistical results.

[0023] A prior set for the examination region is constructed based on the set of regional statistical results, which includes regional identifiers, regional risk prior values, and regional prior adaptation rules.

[0024] Based on the regional group collaborative sensing data set, the trajectory segments of the test vehicles corresponding to the occurrence of historical risk events are extracted, and the trajectory segments are mapped into road segment index sequences according to the pre-configured road segment segmentation rules to generate a set of road segment mapping results.

[0025] Based on the road segment mapping result set, the historical risk event count, historical braking execution count, and historical pedestrian priority trigger count corresponding to each road segment identifier are statistically analyzed. Combined with the road segment unit length and road segment event density, a road segment risk distribution result set is generated. Threshold filtering and connectivity merging are performed to form a candidate set of dangerous road segments.

[0026] Based on the candidate set of dangerous road segments, a prior set of dangerous road segments is constructed, which includes dangerous road segment identifiers, dangerous road segment boundaries, dangerous road segment risk prior values, and dangerous road segment prior adaptation rules.

[0027] The regional group collaborative sensing data set is divided into time buckets according to the pre-configured time period segmentation rules to generate time period bucketed collaborative sensing data sets;

[0028] Based on the time-segmented collaborative sensing data set, the historical risk event count, historical braking execution count, and historical pedestrian priority trigger count corresponding to each time period identifier are statistically analyzed to generate a time period risk distribution result set.

[0029] Construct a time period risk prior set based on the time period risk distribution result set, which includes time period identifier, time period risk prior value and time period prior adaptation rule;

[0030] The prior sets of the examination area, dangerous road sections, and time period risks are merged to generate a prior constraint set for the examination scenario.

[0031] Optionally, the generation of the fused evidence result set specifically includes:

[0032] The multi-sensor collaborative sensing data set is split into window-level collaborative sensing units according to the alignment window. The online status items, data integrity items, measurement stability items and consistency conclusion items bound to the alignment window in the sensor availability status set are written into the window-level collaborative sensing units to generate window-level adjudication input units.

[0033] An evidence proposition space is constructed based on a window-level adjudication input unit, including pedestrian propositions, obstacle propositions, and uncertain propositions. For millimeter-wave radar, depth vision sensor, high-definition camera, ultrasonic radar, and lidar, a set of sensor evidence items consistent with the evidence proposition space is generated respectively.

[0034] Based on the prior constraint set of the examination scenario, perform prior adaptation on the sensor evidence item set to generate a prior adaptation result set within the alignment window;

[0035] Based on the prior adaptation result set and the sensor availability state set, perform evidence discounting on the sensor evidence item set and output the discounted evidence item set.

[0036] Perform evidence theory fusion adjudication on the discounted evidence item set, and output an initial fusion evidence result set, including the fusion evidence strength and corresponding conflict degree for pedestrian propositions, obstacle propositions and uncertain propositions;

[0037] Generate fusion confidence based on the initial fusion evidence result set;

[0038] Pedestrian priority markers are generated based on the initial fusion evidence result set, and together with the fusion confidence results of the corresponding aligned window, they are written into the initial fusion evidence result set to form the final fusion evidence result set.

[0039] Optionally, the generation of the fusion environment representation set specifically includes:

[0040] The final fused evidence result set is read in units of aligned windows, and aggregated by evidence time index to form a window-level fused evidence item set;

[0041] Based on the recursive input evidence unit, a candidate target set is constructed in the signal analysis host. The fusion evidence entries that point to pedestrian propositions and obstacle propositions are mapped to pedestrian candidate target entries and obstacle candidate target entries, respectively. Their corresponding fusion confidence, conflict degree and pedestrian priority mark are bound and written into the candidate target set.

[0042] Based on the candidate target set, target generation gating is performed, and the output evidence passes through the target set;

[0043] Under the constraints of the prior constraint set in the examination scenario, a recursive prior adaptation is performed on the evidence through the target set, and the recursive prior adaptation result is written into the evidence through the target set to form the evidence through the target set prior adaptation result.

[0044] Based on the evidence and the prior adaptation results of the target set, construct the target survival constraint and clutter suppression constraint of the evidence credibility constraint random finite set target set recursively, and output the evidence credibility constraint random finite set target set recursively constraint set;

[0045] Perform the recursion under the constraints of the evidence credibility constraint random finite set target set recursion constraint set, and output the evidence credibility constraint random finite set target set recursion result;

[0046] Based on the recursive results of the random finite set of target sets constrained by the credibility of evidence, the target trajectory is consistently associated and the category is fixed, and the pedestrian target recognition results and obstacle target recognition results are extracted.

[0047] Within the same alignment window, the spatial relationship representation of the target is calculated based on the pedestrian target recognition results and the obstacle target recognition results, and then converged to generate a fused environment representation set.

[0048] Optionally, the generation of the braking execution result specifically includes:

[0049] Read the fusion environment characterization set in units of aligned windows to generate window-level hazard analysis input units;

[0050] Based on the window-level hazard analysis input unit, the pedestrian target recognition results are mapped to pedestrian hazard candidate objects, the obstacle target recognition results are mapped to obstacle hazard candidate objects, and the spatial relationship representation of each hazard candidate object and its corresponding target is bound and written into the hazard analysis candidate object set.

[0051] Based on the candidate hazard analysis set, perform pedestrian priority hazard weight allocation and output the pedestrian priority weight configuration result.

[0052] Under the constraint of pedestrian priority weight configuration results, the risk level calculation is performed on the candidate object set of risk analysis by combining the target spatial relationship representation, and the window-level risk level result is output.

[0053] Based on the window-level hazard level result, generate a trigger evidence retention instruction corresponding to the hazard level result, and bind it to the corresponding window-level hazard level result;

[0054] The window-level hazard level result is sent to the braking control host. Based on the braking level configuration corresponding to the hazard level result, the control current output is configured to drive the brake motor to perform deceleration braking or emergency braking. The braking execution result, the corresponding hazard level result, and the trigger evidence retention instruction are written together into the braking result record set.

[0055] Optionally, the generation of the examination security protection closed loop specifically includes:

[0056] Read the retention time index, retention object type identifier and retention window range identifier from the braking result record set, generate the evidence collection task set, and send it to the vehicle monitoring transmission host;

[0057] The vehicle-mounted monitoring and transmission host calls on the vehicle-mounted video source and vehicle-mounted image source based on the evidence collection task set to generate a multimedia package set of risk events.

[0058] The risk event multimedia package is transmitted to the back-end management system, archived according to vehicle identification and retention timestamp, and risk event retention results are generated.

[0059] The backend management system generates a queue of events to be processed based on the risk event retention results. In the processing interface, the corresponding risk event videos and risk event images are displayed according to the retention index. At the same time, the window-level hazard level results and braking execution results bound to them are displayed. The system receives backend processing actions and writes them into the event processing record set.

[0060] A set of evidence receipts is generated based on the risk event retention results and the event handling record set.

[0061] Based on the evidence receipt set, execute the evidence receipt-driven adaptive discount update rule and output the generated configuration update result;

[0062] The generated configuration update results are written back to the signal analysis host to update the generated configuration of the sensor availability status set and the fusion evidence result set, thus completing the closed loop of exam security protection.

[0063] The driving test safety protection system based on vision and radar fusion according to an embodiment of the present invention includes:

[0064] The environmental perception data acquisition module is used to acquire environmental perception data during locomotive operation, and perform time alignment and signal verification processing to form a multi-source collaborative perception data set, while generating a data availability status set.

[0065] The driving scenario prior construction module is used to construct the prior set of the route area, the prior set of key road sections, and the prior set of time segment risks, and to perform fusion processing to generate the prior constraint set of the driving scenario.

[0066] The fusion adjudication analysis module is used to execute the adjudication processing of evidence fusion rules based on the multi-source collaborative perception data set, the data availability status set, and the prior constraint set of driving scenario, and outputs the fusion evidence result set;

[0067] The target generation and recursion module is used to perform target generation gating processing. Under the constraints of the prior constraint set of the driving scenario, the evidence is recursively processed through the target set to perform evidence credibility constraints, and the fusion environment representation set is extracted.

[0068] The hazard level assessment and alert generation module is used to perform hazard level analysis for pedestrian priority based on the fused environmental characterization set, output hazard level results, and generate corresponding driving alert data and trigger commands;

[0069] The closed-loop update module is used to generate driving risk event retention results based on trigger commands, and perform adaptive discount updates in combination with background processing feedback to complete the closed-loop compilation and optimization of driving prompt data.

[0070] The beneficial effects of this invention are:

[0071] This invention unifies and processes multi-sensor collaborative sensing data sets, sensor availability status sets, and prior constraint sets of examination scenarios through modeling and linkage. At the evidence level, it introduces prior adaptation, evidence discounting, and evidence theory fusion adjudication, enabling fusion confidence, conflict degree, and pedestrian priority marking to stably and controllably participate in the target generation gating and random finite set target recursion process. Thus, even under conditions of inconsistent multi-source sensing timing, local signal anomalies, or complex scene changes, it can still maintain the continuity and consistency of pedestrian and obstacle target recognition, significantly reducing the impact of false detections, missed detections, and short-term jitter on braking decisions, and improving real-time safety protection capabilities during the driving test.

[0072] Meanwhile, this invention constructs a closed-loop security protection mechanism covering perception, decision-making, execution, and post-event correction by synchronizing and storing risk event videos and images in sequence, generating evidence receipt sets through background processing actions, and using adaptive discount update rules driven by evidence receipts. This enables the sensor availability status and fused evidence generation configuration to be continuously updated and rolled back based on real risk events and human processing conclusions, achieving adaptive adjustment under changes in risk characteristics in different examination areas, dangerous road sections, and different time periods, effectively enhancing the system's long-term stability, traceability, and scenario generalization capabilities. Attached Figure Description

[0073] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0074] Figure 1 This is a flowchart of the driving test safety protection method based on vision and radar fusion proposed in this invention;

[0075] Figure 2 This is a schematic diagram illustrating the process of generating the prior constraint set for the driving test scenario in the driving test safety protection method based on vision and radar fusion proposed in this invention.

[0076] Figure 3This is a schematic diagram of the pedestrian priority target generation and fusion environment characterization process of the driving test safety protection method based on vision and radar fusion proposed in this invention. Detailed Implementation

[0077] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0078] refer to Figures 1-3 The safety protection method for driving test part 3 based on vision and radar fusion includes the following steps:

[0079] Acquire environmental perception data and perform time alignment and signal verification to form a multi-sensor collaborative perception data set, and generate a sensor availability status set;

[0080] Based on the multi-sensor collaborative perception data set, a prior set of examination area, a prior set of dangerous road section, and a prior set of time period risk are constructed, and then fused to generate a prior constraint set of examination scenario;

[0081] The multi-sensor collaborative sensing data set, the sensor availability status set, and the examination scenario prior constraint set are input into the signal analysis host, which performs evidence theory fusion adjudication and outputs a fusion evidence result set.

[0082] Based on the fusion evidence result set, the target generation gating is performed to obtain evidence through the target set. Under the constraints of the prior constraint set of the examination scenario, the evidence credibility constraint random finite set target set is recursively calculated. Pedestrian target recognition results, obstacle target recognition results and target spatial relationship representation are extracted to form a fusion environment representation set.

[0083] Based on the fusion of environmental characterization sets, pedestrian priority hazard level analysis is performed, hazard level results and trigger evidence retention instructions are output and sent to the braking control host to implement deceleration braking or emergency braking, forming braking execution results;

[0084] The system sends the trigger evidence retention instruction to the vehicle monitoring and transmission host to generate risk event retention results. Based on the risk event retention results and the background processing actions, it performs adaptive discount updates to complete the closed loop of exam security protection.

[0085] In this embodiment, the generation of the multi-sensor collaborative sensing data set and the sensor availability status set specifically includes:

[0086] Simultaneously acquire millimeter-wave radar data, depth vision sensor data, high-definition camera data, ultrasonic radar data, and lidar data at the test vehicle end to form a raw set of environmental perception data;

[0087] Each frame of environmental sensing data in the original set of environmental sensing data is written with a unified collection timestamp and device identifier.

[0088] Perform time alignment processing on the raw set of environmental perception data, and output a time-aligned data set;

[0089] The time alignment process establishes an alignment window based on a unified fusion period, and uses the rule of minimizing the sum of the absolute deviations between the acquisition timestamps of each sensor frame within the alignment window and the candidate alignment reference time to determine the alignment reference time, and pairs the corresponding frames of each sensor within the alignment window according to the alignment reference time.

[0090] Perform signal verification processing on the time-aligned data set and output a set of signal verification results.

[0091] The signal verification process includes integrity verification, range consistency verification, and mutation verification. The mutation verification uses the absolute value of the difference between the verification feature values ​​of the same sensor in adjacent alignment windows as the mutation degree, and writes the mutation degree, integrity verification result, and range consistency verification result together into the signal verification result set.

[0092] Based on the signal verification result set, a corresponding sensor availability status record is generated for each sensor in each alignment window, and a sensor availability status set is constructed, which includes online status items, data integrity items, measurement stability items and consistency conclusion items.

[0093] Based on the sensor availability status set, the time-aligned data set is subjected to availability filtering and back-off encapsulation to obtain a multi-sensor collaborative sensing data set.

[0094] In this embodiment, the generation of the prior constraint set for the examination scenario specifically includes:

[0095] The location and driving status information of the test vehicle are obtained based on the multi-sensor collaborative perception data set, and spatial matching is performed with the pre-configured test area boundary data to generate a set of test vehicle area affiliation results.

[0096] The examination area boundary data describes the spatial boundaries and area identifiers of each examination room and its respective area, and is written during system deployment;

[0097] Based on the test vehicle area attribution result set, the multi-sensor collaborative perception data set is grouped by region to generate a region-grouped collaborative perception data set;

[0098] Using the region identifier as an index, count the historical risk events, historical braking execution counts, and historical pedestrian priority trigger counts within the region, and output the set of regional statistical results.

[0099] A prior set for the examination region is constructed based on the set of regional statistical results, which includes regional identifiers, regional risk prior values, and regional prior adaptation rules.

[0100] The regional identifiers are derived from the examination area boundary data pre-configured during the system deployment phase. They describe the spatial range of each Subject 3 examination room and its corresponding examination area, and assign a unique regional identifier. The regional risk prior value is generated by performing normalization processing on the regional statistical results to eliminate the influence of differences in the number of examinations in different regions. It is generated by performing weighted summation according to the pre-configured prior fusion weights. For each regional identifier, the regional prior adaptation rule set reads the corresponding regional risk prior value and maps it to the pre-configured risk level interval to determine the risk level status of the region. The corresponding regional prior adaptation rule set is selected from the predefined rule generation mapping table, including evidence discount adaptation rules, pedestrian priority enhancement rules, target generation gating bias rules, and conflict suppression bias rules.

[0101] Based on the regional group collaborative sensing data set, the trajectory segments of the test vehicles corresponding to the occurrence of historical risk events are extracted, and the trajectory segments are mapped into road segment index sequences according to the pre-configured road segment segmentation rules to generate a set of road segment mapping results.

[0102] The road segmentation rules are used to divide the test route into non-overlapping road segment units and assign a road segment identifier to each road segment unit.

[0103] Based on the road segment mapping result set, the historical risk event count, historical braking execution count, and historical pedestrian priority trigger count corresponding to each road segment identifier are statistically analyzed. Combined with the road segment unit length and road segment event density, a road segment risk distribution result set is generated. Threshold filtering and connectivity merging are performed to form a candidate set of dangerous road segments.

[0104] Based on the candidate set of dangerous road segments, a prior set of dangerous road segments is constructed, which includes dangerous road segment identifiers, dangerous road segment boundaries, dangerous road segment risk prior values, and dangerous road segment prior adaptation rules.

[0105] The dangerous road segment identifiers are derived from the road segmentation rules and road segment identifier allocation table pre-configured during the system deployment phase. Each road segment unit is assigned a unique corresponding dangerous road segment identifier. The dangerous road segment boundary is determined by the road segmentation rules and stored in the form of spatial boundary description information, which limits the spatial range of the dangerous road segment unit and serves as the basis for road segment mapping between risk events and trajectory segments. The prior risk value of the dangerous road segment is indexed by the dangerous road segment identifier. The impact of differences in road segment length and number of passes is eliminated by performing normalization processing on the road segment risk distribution result set. The prior fusion weights are used to generate the value. For each dangerous road segment identifier, the prior adaptation rule set reads the corresponding dangerous road segment risk prior value, maps it to the pre-configured risk level interval to determine the corresponding risk level state, and selects the corresponding dangerous road segment prior adaptation rule set from the predefined rule generation mapping table, including the target generation gating bias rule and the observation update suppression bias rule.

[0106] The regional group collaborative sensing data set is divided into time buckets according to the pre-configured time period segmentation rules to generate time period bucketed collaborative sensing data sets;

[0107] The time-segmentation rule divides the exam time within a day into non-overlapping time-segment units and assigns a time-segment identifier to each time-segment unit.

[0108] Based on the time-segmented collaborative sensing data set, the historical risk event count, historical braking execution count, and historical pedestrian priority trigger count corresponding to each time period identifier are statistically analyzed to generate a time period risk distribution result set.

[0109] Construct a time period risk prior set based on the time period risk distribution result set, which includes time period identifier, time period risk prior value and time period prior adaptation rule;

[0110] The time period identifier is derived from the time period segmentation rules and time period identifier allocation table pre-configured during the system deployment phase. A unique time period identifier is assigned to each time period unit. The time period risk prior value is indexed by the time period identifier. The impact of the difference in the number of exams in different time periods is eliminated by performing normalization processing on the time period risk distribution result set. The time period risk prior value is generated by performing weighted summation according to the pre-configured prior fusion weight. For each time period identifier, the time period prior adaptation rule set reads the corresponding time period risk prior value and maps it to the pre-configured risk level range to determine the corresponding risk level status. The corresponding time period prior adaptation rule set is selected from the predefined rule generation mapping table, including the evidence discount bias rule and the conflict degree suppression bias rule.

[0111] The prior sets of the examination area, dangerous road sections, and time period risks are merged to generate a prior constraint set for the examination scenario.

[0112] In this embodiment, the generation of the fused evidence result set specifically includes:

[0113] The multi-sensor collaborative sensing data set is split into window-level collaborative sensing units according to the alignment window. The online status items, data integrity items, measurement stability items and consistency conclusion items bound to the alignment window in the sensor availability status set are written into the window-level collaborative sensing units to generate window-level adjudication input units.

[0114] An evidence proposition space is constructed based on a window-level adjudication input unit, including pedestrian propositions, obstacle propositions, and uncertain propositions. For millimeter-wave radar, depth vision sensor, high-definition camera, ultrasonic radar, and lidar, a set of sensor evidence items consistent with the evidence proposition space is generated respectively.

[0115] The evidence proposition space is assigned proposition labels and mutual exclusion constraints to pedestrian propositions, obstacle propositions, and uncertain propositions.

[0116] Each evidence entry in the sensor evidence entry set includes an evidence source identifier, proposition orientation, evidence strength description, and evidence time index;

[0117] Based on the prior constraint set of the examination scenario, perform prior adaptation on the sensor evidence item set to generate a prior adaptation result set within the alignment window;

[0118] The prior adaptation includes reading the prior adaptation rules of the region according to the region identifier, reading the prior adaptation rules of the dangerous road segment according to the dangerous road segment identifier, and reading the prior adaptation rules of the time period according to the time period identifier.

[0119] The prior fitting result set limits the fitting values ​​of evidence discount fitting, pedestrian priority enhancement, target generation gating bias, conflict suppression bias, observation update suppression bias, and evidence discount bias within this alignment window.

[0120] Based on the prior adaptation result set and the sensor availability state set, perform evidence discounting on the sensor evidence item set and output the discounted evidence item set.

[0121] The evidence discounting process is implemented by multiplying the discount coefficient by the evidence strength before discounting for the corresponding proposition to obtain the evidence strength after discounting for the corresponding proposition. The discount coefficient is jointly determined by the online status item, data integrity item, measurement stability item and consistency conclusion item, and is subject to evidence discounting adaptation and evidence discounting bias constraints.

[0122] Perform evidence theory fusion adjudication on the discounted evidence item set, and output an initial fusion evidence result set, including the fusion evidence strength and corresponding conflict degree for pedestrian propositions, obstacle propositions and uncertain propositions;

[0123] The evidence theory fusion adjudication adopts the rule of accumulating the discounted evidence strength from different evidence sources according to the proposition intersection being consistent and accumulating it as the conflict degree when the proposition intersection is empty, and normalizing and suppressing the conflict degree. It outputs the fused evidence strength for pedestrian propositions, obstacle propositions and uncertain propositions, and simultaneously outputs the conflict degree bound to the fused evidence strength.

[0124] Generate fusion confidence based on the initial fusion evidence result set;

[0125] The fusion confidence score is determined by the propositional orientation corresponding to the maximum fusion strength in the initial fusion evidence result set and its evidence strength descriptor, and is output as a fusion confidence score result subject to conflict suppression bias constraint.

[0126] Based on the initial fusion evidence result set, a pedestrian priority marker is generated and written together with the fusion confidence result of the corresponding aligned window into the initial fusion evidence result set to form the final fusion evidence result set;

[0127] The pedestrian priority label is obtained by jointly deciding the pedestrian proposition fusion evidence strength in the initial fusion evidence result set, the pedestrian priority enhancement rule in the regional prior adaptation rule, and the conflict degree suppression bias in the time period prior adaptation rule.

[0128] In this embodiment, the generation of the fusion environment representation set specifically includes:

[0129] The final fused evidence result set is read in units of aligned windows, and aggregated by evidence time index to form a window-level fused evidence item set;

[0130] Each fused evidence entry in the window-level fused evidence entry set includes proposition orientation, fused evidence strength, conflict degree, fused confidence degree and pedestrian priority mark, generating a recursive input evidence unit;

[0131] Based on the recursive input evidence unit, a candidate target set is constructed in the signal analysis host. The fusion evidence entries that point to pedestrian propositions and obstacle propositions are mapped to pedestrian candidate target entries and obstacle candidate target entries, respectively. Their corresponding fusion confidence, conflict degree and pedestrian priority mark are bound and written into the candidate target set.

[0132] Based on the candidate target set, target generation gating is performed, and the output evidence passes through the target set;

[0133] The target generation gating is implemented using gating pass determination rules. The fusion confidence is used as the main criterion for evidence pass determination, the conflict degree is used as the gating suppression criterion, and the pedestrian priority mark is used as the gating relaxation criterion. When the fusion confidence meets the evidence pass condition, the corresponding candidate target item enters the evidence pass target set. When the conflict degree meets the suppression condition, the corresponding candidate target item is removed from the evidence pass target set or downgraded to an uncertain candidate target item. When the pedestrian priority mark meets the trigger condition, the corresponding pedestrian candidate target item enters the evidence pass target set with priority under the same fusion confidence.

[0134] Under the constraints of the prior constraint set in the examination scenario, a recursive prior adaptation is performed on the evidence through the target set, and the recursive prior adaptation result is written into the evidence through the target set to form the evidence through the target set prior adaptation result.

[0135] The recursive prior adaptation includes reading the target to generate gating bias rules and conflict suppression bias rules according to the region identifier, reading the observation update suppression bias rules according to the dangerous road segment identifier, and reading the conflict degree suppression bias rules according to the time period identifier.

[0136] Based on the evidence and the prior adaptation results of the target set, construct the target survival constraint and clutter suppression constraint of the evidence credibility constraint random finite set target set recursively, and output the evidence credibility constraint random finite set target set recursively constraint set;

[0137] The target survival constraint uses fusion confidence as the basis for survival enhancement and conflict degree as the basis for survival suppression, and is subject to conflict suppression bias rules and conflict degree suppression bias rules. The clutter suppression constraint uses conflict degree as the basis for clutter enhancement and is subject to observation update suppression bias rules.

[0138] Perform the recursion under the constraints of the evidence credibility constraint random finite set target set recursion constraint set, and output the evidence credibility constraint random finite set target set recursion result;

[0139] The recursion includes target prediction, target survival screening, observation update suppression, and clutter suppression. Observation update suppression uses conflict degree as the suppression entry point and is restricted by the observation update suppression bias rule. Target survival screening uses fusion confidence degree as the enhancement entry point and is restricted by the target generation gating bias rule.

[0140] Based on the recursive results of the random finite set of target sets constrained by the credibility of evidence, the target trajectory is consistently associated and the category is fixed, and the pedestrian target recognition results and obstacle target recognition results are extracted.

[0141] The pedestrian target recognition result is generated when the survival and association consistency of the pedestrian candidate target entries in the recursive window meet the solidification condition; the obstacle target recognition result is generated when the survival and association consistency of the obstacle candidate target entries in the recursive window meet the solidification condition.

[0142] Within the same alignment window, the spatial relationship representation of the target is calculated based on the pedestrian target recognition results and the obstacle target recognition results, and then converged to generate a fused environment representation set.

[0143] The target spatial relationship representation includes relative distance relationship, relative orientation relationship and relative proximity trend relationship. Each environmental representation record in the fused environmental representation set includes pedestrian target recognition result, obstacle target recognition result and corresponding target spatial relationship representation.

[0144] In this embodiment, the generation of the braking execution result specifically includes:

[0145] Read the fusion environment characterization set in units of aligned windows to generate window-level hazard analysis input units;

[0146] Based on the window-level hazard analysis input unit, the pedestrian target recognition results are mapped to pedestrian hazard candidate objects, the obstacle target recognition results are mapped to obstacle hazard candidate objects, and the spatial relationship representation of each hazard candidate object and its corresponding target is bound and written into the hazard analysis candidate object set.

[0147] Based on the candidate object set for hazard analysis, perform pedestrian priority hazard weight allocation and output the pedestrian priority weight configuration result;

[0148] The pedestrian priority hazard weight allocation assigns a higher base weight to pedestrian hazard candidate objects than to obstacle hazard candidate objects in the hazard analysis candidate object set, and applies a suppression weight to obstacle hazard candidate objects within the same alignment window when there is a pedestrian target recognition result.

[0149] Under the constraint of pedestrian priority weight configuration results, the risk level calculation is performed on the candidate object set of risk analysis by combining the target spatial relationship representation, and the window-level risk level result is output.

[0150] The risk level calculation is based on a comprehensive evaluation of relative distance relationship, relative proximity trend relationship and relative orientation relationship. When there is a pedestrian target identification result and the relative proximity trend relationship meets the proximity condition, the risk level of the corresponding pedestrian risk candidate is increased.

[0151] Based on the window-level hazard level result, generate a trigger evidence retention instruction corresponding to the hazard level result, and bind it to the corresponding window-level hazard level result;

[0152] The trigger evidence retention instruction includes a retention trigger identifier, a retention time index, a retention object type identifier, and a retention window range identifier;

[0153] The window-level hazard level result is sent to the braking control host. Based on the braking level configuration corresponding to the hazard level result, the control current output is configured to drive the brake motor to perform deceleration braking or emergency braking. The braking execution result, the corresponding hazard level result, and the trigger evidence retention instruction are written together into the braking result record set.

[0154] In this embodiment, the generation of the examination security protection closed loop specifically includes:

[0155] Read the retention time index, retention object type identifier and retention window range identifier from the braking result record set, generate the evidence collection task set, and send it to the vehicle monitoring transmission host;

[0156] The vehicle-mounted monitoring and transmission host calls on the vehicle-mounted video source and vehicle-mounted image source based on the evidence collection task set to generate a multimedia package set of risk events.

[0157] The risk event multimedia package collects risk event videos and risk event images that are time-aligned with the braking execution results. It writes a unified retention timestamp, vehicle identifier, alignment window index, and braking level identifier into the risk event videos and images according to the retention time index.

[0158] The risk event multimedia package is transmitted to the back-end management system, archived according to vehicle identification and retention timestamp, and risk event retention results are generated.

[0159] The risk event retention results include retention index identifier, retention object type identifier, retention window range identifier, corresponding window-level hazard level result, corresponding braking execution result, risk event video index, and risk event image index;

[0160] The backend management system generates a queue of events to be processed based on the risk event retention results. In the processing interface, the corresponding risk event videos and risk event images are displayed according to the retention index. At the same time, the window-level hazard level results and braking execution results bound to them are displayed. The system receives backend processing actions and writes them into the event processing record set.

[0161] The background processing action includes a processing status identifier, a processing conclusion identifier, an evidence consistency conclusion identifier, and a disposal remarks information;

[0162] A set of evidence receipts is generated based on the risk event retention results and the event handling record set.

[0163] The evidence receipt set includes a receipt index identifier, a retention index identifier, a processing status identifier, a processing conclusion identifier, an evidence consistency conclusion identifier, a corresponding window-level hazard level result, a corresponding braking execution result, a conflict source identifier, a sensor availability receipt item, and a discount update suggestion item. The sensor availability receipt item is used to indicate the receipt conclusions of the online status item, data integrity item, measurement stability item, and consistency conclusion item. The discount update suggestion item is used to indicate the set of update actions of the adaptive discount update rule driven by the evidence receipt.

[0164] Based on the evidence receipt set, execute the evidence receipt-driven adaptive discount update rule and output the generated configuration update result;

[0165] The adaptive discount update rule limits the discount update action set to increase, decrease, freeze, and rollback. The target is limited to the generation configuration of the sensor availability status set and the fusion evidence result set. When the processing conclusion identifier represents a valid risk and the evidence consistency conclusion identifier represents consistency, the corresponding discount update suggestion is mapped to increase and apply to the evidence discount bias and target generation gating bias rules of the pedestrian priority enhancement rule. When the processing conclusion identifier represents an invalid risk and the evidence consistency conclusion identifier represents consistency, the corresponding discount update suggestion is mapped to decrease and apply to the generation configuration of the conflict suppression bias and observation update suppression bias. When the processing status identifier represents no processing, the corresponding discount update suggestion is mapped to freeze and keep the existing generation configuration unchanged. When the evidence consistency conclusion identifier represents inconsistency, the corresponding discount update suggestion is mapped to rollback and roll back the generation configuration of the corresponding alignment window to the pre-configured safe baseline version.

[0166] The generated configuration update results are written back to the signal analysis host to update the generated configuration of the sensor availability status set and the fusion evidence result set, thus completing the closed loop of exam security protection.

[0167] The driving test safety protection system based on vision and radar fusion includes:

[0168] The environmental perception data acquisition module is used to acquire environmental perception data during locomotive operation, and perform time alignment and signal verification processing to form a multi-source collaborative perception data set, while generating a data availability status set.

[0169] The driving scenario prior construction module is used to construct the prior set of the route area, the prior set of key road sections, and the prior set of time segment risks, and to perform fusion processing to generate the prior constraint set of the driving scenario.

[0170] The fusion adjudication analysis module is used to execute the adjudication processing of evidence fusion rules based on the multi-source collaborative perception data set, the data availability status set, and the prior constraint set of driving scenario, and outputs the fusion evidence result set;

[0171] The target generation and recursion module is used to perform target generation gating processing. Under the constraints of the prior constraint set of the driving scenario, the evidence is recursively processed through the target set to perform evidence credibility constraints, and the fusion environment representation set is extracted.

[0172] The hazard level assessment and alert generation module is used to perform hazard level analysis for pedestrian priority based on the fused environmental characterization set, output hazard level results, and generate corresponding driving alert data and trigger commands;

[0173] The closed-loop update module is used to generate driving risk event retention results based on trigger commands, and perform adaptive discount updates in combination with background processing feedback to complete the closed-loop compilation and optimization of driving prompt data.

[0174] Example 1:

[0175] To verify the feasibility of this invention in practice, it was applied to the actual testing environment of multiple driving test (Part 3) examination sites in a certain city. In this scenario, the test roads are complex, including straight sections, U-turn sections, parking areas, and public roads frequently crossed by pedestrians and non-motorized vehicles. Traditional testing methods rely on onboard safety personnel, which has long suffered from problems such as manual prompting, substitute braking, and difficulty in leaving traces of supervision, directly affecting the fairness of the test and road safety. After deployment in this scenario, this invention uses onboard multi-source sensing devices to continuously sense and record the entire testing process. Without relying on onboard safety personnel, it can identify and assess changes in the surrounding environment of the test vehicle in real time, effectively solving the problems of difficult-to-eliminate human interference and the lack of unified standards for risk handling during the testing process. This provides a technical path that restores the testing process to objectivity, traceability, and supervision.

[0176] In practical applications, after the test vehicle enters the test route, the system continuously collects environmental perception data from millimeter-wave radar, visual sensors, and ultrasonic radar, and forms stable collaborative perception results through time alignment and signal verification. The system dynamically generates prior constraints for the test scenario by combining pre-configured test area information, historical high-risk road section distribution, and risk characteristics at different times, and performs targeted constraint modeling for the current test environment. During the journey, when the system detects pedestrians crossing, vehicles slowing down, deviations from parking on the side of the road, or abnormal distances between the vehicle and barriers, it prioritizes pedestrian and obstacle risks through a fusion of evidence adjudication and target deduction mechanisms, and automatically triggers braking control when necessary. Simultaneously, the system automatically retains the preceding and following time-series video and image data corresponding to the braking behavior and uploads them to the backend management system, achieving full-process traceability and post-event verification, ensuring both test order and road safety are simultaneously guaranteed.

[0177] From an operational perspective, the system maintained stable operation over a considerable period. Test vehicles did not experience frequent accidental braking under normal driving conditions. In high-risk scenarios such as U-turns, following other vehicles, and pedestrian crossings, the system was able to intervene promptly and handle situations safely. The backend system continuously aggregates and stores information on risk events, statistically analyzing triggering road sections, time periods, and types, providing a reliable basis for optimizing test routes and managing on-site order. By introducing the method of this invention, human intervention during the testing process was significantly reduced, testing efficiency was significantly improved, test vehicle turnaround capacity was enhanced, and the frequency of test-related traffic safety incidents decreased significantly. The system's stable operation and traceable management also significantly reduced candidates' concerns about the fairness of the test, improved the management efficiency and social acceptance of test organizers, and fully demonstrated the technical value and application effect of this invention in real-world testing scenarios.

[0178] Table 1. Performance Comparison of Vision and Radar Fusion-Based Safety Protection Method for Subject 3 Examination with Traditional Methods

[0179] Comparison indicators Traditional safety officer on-vehicle testing model Examination mode using the method of this invention Number of exam venues covered 10 10 Number of vehicles put into the test 224 vehicles 224 vehicles Total number of test takers More than 56,000 people More than 56,000 people Total number of braking triggers 623 times 556 times Braking trigger rate 1.10% 0.98% Average number of exam accidents per month 19 cases 12 cases Average number of people that can be processed per exam Limited by the number of safety personnel All vehicles operating at full capacity Exam pass rate 63% 74% Manual intervention prompts record exist No occurrence Risk event traceability rate Partial retention Full retention Number of exam complaints and inquiries More 56 times Verifiability of the examination process Relying on manual explanation Videos and images can be directly verified.

[0180] As shown in Table 1, under the premise of consistent number of examinees, vehicles, and examination venue size, the safety, stability, and fairness of the examination process are significantly improved by adopting the method of this invention. Firstly, regarding the braking trigger rate, the system triggered the brakes a total of 556 times during the examination, lower than the 623 times in the traditional mode, reducing the braking trigger rate from 1.10% to 0.98%. This result demonstrates that this invention, through multi-sensor collaborative perception, evidence theory fusion, and prior constraints on the examination scenario, effectively reduces misjudgments and unnecessary braking triggers, making braking behavior more focused on real high-risk scenarios and avoiding redundant triggers caused by premature human intervention or experience-based braking.

[0181] In terms of traffic safety, the decrease in the number of accidents is particularly significant. Under the traditional testing model, an average of 19 accidents occur per month during the testing process, while the number drops to 12 per month after adopting the method of this invention. This improvement is not due to a reduction in testing intensity, but rather achieved under conditions of full-load operation of the testing vehicles. This demonstrates that the pedestrian priority hazard level analysis, target generation gating, and evidence credibility constraint target recursion mechanism introduced in this invention can identify pedestrian and obstacle risks in advance in complex testing road conditions and intervene with braking in a timely manner at critical moments, fundamentally reducing the probability of accidents.

[0182] In terms of examination efficiency and organizational capacity, the adoption of this invention eliminates the need for safety officers to accompany examinees in the vehicle, allowing for continuous operation of the examination vehicles and significantly alleviating the backlog caused by insufficient personnel. Simultaneously, the pass rate increased from 63% to 74%, reflecting that examinees can complete the examination independently without human intervention, which is more conducive to demonstrating their actual driving abilities. This improvement is closely related to the stable generation of integrated environmental representations and the unified hazard level determination criteria in this invention, making the examination process more consistent and predictable.

[0183] Furthermore, from the perspectives of exam management and public feedback, the method of this invention enables the automatic retention of videos and images of risk events throughout the entire exam process. The backend can directly review the retained results, significantly reducing the number of exam complaints and inquiries. The traceability of risk events has been improved from partial retention to full retention, effectively solving the problems of difficulty in defining responsibility and reliance on manual explanations for post-exam verification under the traditional model. In summary, the data shown in Table 1 fully demonstrates that this invention achieves a comprehensive technical effect of significantly improving security, increasing exam efficiency, and enhancing exam credibility without increasing exam costs.

[0184] 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 safety protection method for driving test part 3 based on vision and radar fusion, characterized in that, Includes the following steps: Acquire environmental perception data and perform time alignment and signal verification to form a multi-sensor collaborative perception data set, and generate a sensor availability status set; Based on the multi-sensor collaborative perception data set, a prior set of examination area, a prior set of dangerous road section, and a prior set of time period risk are constructed, and then fused to generate a prior constraint set of examination scenario; The multi-sensor collaborative sensing data set, the sensor availability status set, and the examination scenario prior constraint set are input into the signal analysis host, which performs evidence theory fusion adjudication and outputs a fusion evidence result set. Based on the fusion evidence result set, the target generation gating is performed to obtain evidence through the target set. Under the constraints of the prior constraint set of the examination scenario, the evidence credibility constraint random finite set target set is recursively calculated. Pedestrian target recognition results, obstacle target recognition results and target spatial relationship representation are extracted to form a fusion environment representation set. Based on the fusion of environmental characterization sets, pedestrian priority hazard level analysis is performed, hazard level results and trigger evidence retention instructions are output and sent to the braking control host to implement deceleration braking or emergency braking, forming braking execution results; The system sends the trigger evidence retention instruction to the vehicle monitoring and transmission host to generate risk event retention results. Based on the risk event retention results and the background processing actions, it performs adaptive discount updates to complete the closed loop of exam security protection.

2. The method for ensuring safety during the driving test (subject 3) based on vision and radar fusion as described in claim 1, characterized in that, The generation of the multi-sensor collaborative sensing data set and the sensor availability status set specifically includes: Simultaneously acquire millimeter-wave radar data, depth vision sensor data, high-definition camera data, ultrasonic radar data, and lidar data at the test vehicle end to form a raw set of environmental perception data; Perform time alignment processing on the raw set of environmental perception data, and output a time-aligned data set; Perform signal verification processing on the time-aligned data set and output a set of signal verification results. Based on the signal verification result set, a corresponding sensor availability status record is generated for each sensor in each alignment window, and a sensor availability status set is constructed, which includes online status items, data integrity items, measurement stability items and consistency conclusion items. Based on the sensor availability status set, the time-aligned data set is subjected to availability filtering and back-off encapsulation to obtain a multi-sensor collaborative sensing data set.

3. The method for ensuring safety during the driving test (subject 3) based on vision and radar fusion as described in claim 1, characterized in that, The generation of the prior constraint set for the examination scenario specifically includes: The location and driving status information of the test vehicle are obtained based on the multi-sensor collaborative perception data set, and spatial matching is performed with the pre-configured test area boundary data to generate a set of test vehicle area affiliation results. Based on the test vehicle area attribution result set, the multi-sensor collaborative perception data set is grouped by region to generate a region-grouped collaborative perception data set; Using the region identifier as an index, count the historical risk events, historical braking executions, and historical pedestrian priority triggers within the region, and output the set of regional statistical results. A prior set for the examination region is constructed based on the set of regional statistical results, which includes regional identifiers, regional risk prior values, and regional prior adaptation rules. Based on the regional group collaborative sensing data set, the trajectory segments of the test vehicles corresponding to the occurrence of historical risk events are extracted, and the trajectory segments are mapped into road segment index sequences according to the pre-configured road segment segmentation rules to generate a set of road segment mapping results. Based on the road segment mapping result set, the historical risk event count, historical braking execution count, and historical pedestrian priority trigger count corresponding to each road segment identifier are statistically analyzed. Combined with the road segment unit length and road segment event density, a road segment risk distribution result set is generated. Threshold filtering and connectivity merging are performed to form a candidate set of dangerous road segments. Based on the candidate set of dangerous road segments, a prior set of dangerous road segments is constructed, which includes dangerous road segment identifiers, dangerous road segment boundaries, dangerous road segment risk prior values, and dangerous road segment prior adaptation rules. The regional group collaborative sensing data set is divided into time buckets according to the pre-configured time period segmentation rules to generate time period bucketed collaborative sensing data sets; Based on the time-segmented collaborative sensing data set, the historical risk event count, historical braking execution count, and historical pedestrian priority trigger count corresponding to each time period identifier are statistically analyzed to generate a time period risk distribution result set. A time period risk prior set is constructed based on the time period risk distribution result set, which includes time period identifiers, time period risk prior values ​​and time period prior adaptation rules; The prior sets of the examination area, dangerous road sections, and time period risks are merged to generate a prior constraint set for the examination scenario.

4. The method for ensuring safety during the driving test (subject 3) based on vision and radar fusion as described in claim 1, characterized in that, The generation of the fused evidence result set specifically includes: The multi-sensor collaborative sensing data set is split into window-level collaborative sensing units according to the alignment window. The online status items, data integrity items, measurement stability items and consistency conclusion items bound to the alignment window in the sensor availability status set are written into the window-level collaborative sensing units to generate window-level adjudication input units. Based on the window-level adjudication input unit, an evidence proposition space is constructed, including pedestrian propositions, obstacle propositions and uncertain propositions. For millimeter-wave radar, depth vision sensor, high-definition camera, ultrasonic radar and lidar, a set of sensor evidence items consistent with the evidence proposition space is generated respectively. Based on the prior constraint set of the examination scenario, perform prior adaptation on the sensor evidence item set to generate a prior adaptation result set within the alignment window; Based on the prior adaptation result set and the sensor availability state set, perform evidence discounting on the sensor evidence item set and output the discounted evidence item set. Perform evidence theory fusion adjudication on the discounted evidence item set, and output an initial fusion evidence result set, including the fusion evidence strength and corresponding conflict degree for pedestrian propositions, obstacle propositions and uncertain propositions; Generate fusion confidence based on the initial fusion evidence result set; Pedestrian priority markers are generated based on the initial fusion evidence result set, and together with the fusion confidence results of the corresponding aligned window, they are written into the initial fusion evidence result set to form the final fusion evidence result set.

5. The method for ensuring safety during the driving test (subject 3) based on vision and radar fusion as described in claim 1, characterized in that, The generation of the fusion environment representation set specifically includes: The final fused evidence result set is read in units of aligned windows, and aggregated by evidence time index to form a window-level fused evidence item set; Based on the recursive input evidence unit, a candidate target set is constructed in the signal analysis host. The fusion evidence entries that point to pedestrian propositions and obstacle propositions are mapped to pedestrian candidate target entries and obstacle candidate target entries, respectively. Their corresponding fusion confidence, conflict degree and pedestrian priority mark are bound and written into the candidate target set. Based on the candidate target set, target generation gating is performed, and the output evidence passes through the target set; Under the constraints of the prior constraint set in the examination scenario, a recursive prior adaptation is performed on the evidence through the target set, and the recursive prior adaptation result is written into the evidence through the target set to form the evidence through the target set prior adaptation result. Based on the evidence and the prior adaptation results of the target set, construct the target survival constraint and clutter suppression constraint of the evidence credibility constraint random finite set target set recursively, and output the evidence credibility constraint random finite set target set recursively constraint set; Perform the recursion under the constraints of the evidence credibility constraint random finite set target set recursion constraint set, and output the evidence credibility constraint random finite set target set recursion result; Based on the recursive results of the random finite set of target sets constrained by the credibility of evidence, the target trajectory is consistently associated and the category is fixed, and the pedestrian target recognition results and obstacle target recognition results are extracted. Within the same alignment window, the spatial relationship representation of the target is calculated based on the pedestrian target recognition results and the obstacle target recognition results, and then converged to generate a fused environment representation set.

6. The method for ensuring safety during the driving test (subject 3) based on vision and radar fusion as described in claim 1, characterized in that, The generation of the braking execution result specifically includes: Read the fusion environment characterization set in units of aligned windows to generate window-level hazard analysis input units; Based on the window-level hazard analysis input unit, the pedestrian target recognition results are mapped to pedestrian hazard candidate objects, the obstacle target recognition results are mapped to obstacle hazard candidate objects, and the spatial relationship representation of each hazard candidate object and its corresponding target is bound and written into the hazard analysis candidate object set. Based on the candidate hazard analysis set, perform pedestrian priority hazard weight allocation and output the pedestrian priority weight configuration result. Under the constraint of pedestrian priority weight configuration results, the risk level calculation is performed on the candidate object set of risk analysis by combining the target spatial relationship representation, and the window-level risk level result is output. Based on the window-level hazard level result, generate a trigger evidence retention instruction corresponding to the hazard level result, and bind it to the corresponding window-level hazard level result; The window-level hazard level result is sent to the braking control host. Based on the braking level configuration corresponding to the hazard level result, the control current output is configured to drive the brake motor to perform deceleration braking or emergency braking. The braking execution result, the corresponding hazard level result, and the trigger evidence retention instruction are written together into the braking result record set.

7. The method for ensuring safety during the driving test (subject 3) based on vision and radar fusion as described in claim 1, characterized in that, The generation of the examination security protection closed loop specifically includes: Read the retention time index, retention object type identifier and retention window range identifier from the braking result record set, generate the evidence collection task set, and send it to the vehicle monitoring transmission host; The vehicle-mounted monitoring and transmission host calls on the vehicle-mounted video source and vehicle-mounted image source based on the evidence collection task set to generate a multimedia package set of risk events. The risk event multimedia package is transmitted to the back-end management system, archived according to vehicle identification and retention timestamp, and risk event retention results are generated. The backend management system generates a queue of events to be processed based on the risk event retention results. In the processing interface, the corresponding risk event videos and risk event images are displayed according to the retention index. At the same time, the window-level hazard level results and braking execution results bound to them are displayed. The system receives backend processing actions and writes them into the event processing record set. A set of evidence receipts is generated based on the risk event retention results and the event handling record set. Based on the evidence receipt set, execute the evidence receipt-driven adaptive discount update rule and output the generated configuration update result; The generated configuration update results are written back to the signal analysis host to update the generated configuration of the sensor availability status set and the fusion evidence result set, thus completing the closed loop of exam security protection.

8. A driving test safety protection system based on vision and radar fusion, implementing the driving test safety protection method based on vision and radar fusion as described in any one of claims 1 to 7, characterized in that, include: The environmental perception data acquisition module is used to acquire environmental perception data during locomotive operation, and perform time alignment and signal verification processing to form a multi-source collaborative perception data set, while generating a data availability status set. The driving scenario prior construction module is used to construct the prior set of the route area, the prior set of key road sections, and the prior set of time segment risks, and to perform fusion processing to generate the prior constraint set of the driving scenario. The fusion adjudication analysis module is used to execute the adjudication processing of evidence fusion rules based on the multi-source collaborative perception data set, the data availability status set, and the prior constraint set of driving scenario, and outputs the fusion evidence result set; The target generation and recursion module is used to perform target generation gating processing. Under the constraints of the prior constraint set of the driving scenario, the evidence is recursively processed through the target set to perform evidence credibility constraints, and the fusion environment representation set is extracted. The hazard level assessment and alert generation module is used to perform hazard level analysis for pedestrian priority based on the fused environmental characterization set, output hazard level results, and generate corresponding driving alert data and trigger commands; The closed-loop update module is used to generate driving risk event retention results based on trigger commands, and perform adaptive discount updates in combination with background processing feedback to complete the closed-loop compilation and optimization of driving prompt data.

Images