{"id":"6d842b39-b45a-4166-a64b-249cd48e9929","arxiv_id":"2507.18160","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A drone subsystem combining YOLOv11 detection, dlib face recognition, and pose-based distance estimation can find, recognize, and follow a specific person in real time.","lead":"The paper describes a ROS2-based system that lets a small DJI Tello drone detect people, recognize a target face, and follow that person using YOLOv11, YOLOv11-pose, dlib, and tuned PD controllers. The demonstration works in real time on a handful of known individuals, but the paper provides qualitative results only and omits key fitted parameters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 1's fixed shoulder-to-hip distance calibration is the load-bearing link: all three PD controllers regulate an unvalidated pixel-to-metric mapping, so the claimed 2 m tracking distance is not actually demonstrated.","rationale":"The reader's weakest assumption correctly identifies the distance calibration in Eq. 1 as the link on which all three PD controllers depend. I considered the 14-versus-11 subject inconsistency and the absence of quantitative tracking metrics; these are reporting weaknesses, but the distance calibration is more load-bearing because even perfect detection, recognition, and pose estimation would not yield the claimed behavior if the distance signal is biased. The proposed concrete test directly checks whether the mapping is valid across the population the system would encounter in search and rescue. This does not move the verdict: the paper remains CONDITIONAL because it is a plausible proof of concept, but the central quantitative claim about maintaining a specified physical distance is not yet demonstrated.","tokens_in":6024,"tokens_out":3344,"duration_ms":40314,"concrete_test":"Measure the Eq. 1 distance estimator against ground truth using the published GitHub ROS2 pipeline: place individuals of several heights (e.g., 150, 165, 180, and 195 cm) at known distances from 0.5 to 6 m, in upright, rotated, and forward-leaning postures, and compare the estimated distance to the measured distance. If the median absolute error at the nominal 2 m setpoint exceeds about 0.3 m for typical adult heights, or if the error varies systematically with subject height, then the claimed predefined-distance tracking is not achieved for the general population.","verdict_should_be":"UNCHANGED","load_bearing_attack":"All three PD controllers described in Section 3.1 consume a single distance signal, y = k1*x^2 + k2*x + k3 [cm] (Eq. 1), where x is the shoulder-to-hip pixel distance from YOLOv11-pose and the polynomial is calibrated for an average European male/female height. The paper gives no fitted coefficient values, no calibration protocol, and no comparison of the estimate against measured distances. Because the tracked individual is arbitrary, a shorter or taller person maps the same physical distance to a different pixel span, so the controllers regulate an estimate that can be systematically biased. The text itself acknowledges that leaning distorts the shoulder-to-hip span and only mitigates this by restricting drone motion on rapid distance changes. Thus the central claim that the UAV 'maintains a specified distance, set to 2 meters' rests on an unvalidated geometric assumption, with no ground-truth error statistics reported. The real-time integration is credible, but the metric-distance component of the claim is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an autonomous UAV subsystem for search and rescue built on a DJI Tello equipped with a Jetson Xavier NX and running ROS2. The system uses YOLOv11 for people detection, dlib for face recognition, and YOLOv11-pose for body keypoint extraction, and it computes a monocular distance estimate from the shoulder-to-hip pixel span. Three PD controllers, tuned through data-driven system identification, regulate forward/backward, vertical, and yaw motion, with a target tracking distance set to 2 m. The authors report qualitative real-time indoor and outdoor experiments and state that the complete solution, videos, and evaluations are available on GitHub.","tokens_in":6359,"tokens_out":3742,"duration_ms":37271,"significance":"If properly validated, the system would be a useful proof of concept for a compact, low-cost UAV performing the full detect-recognize-track loop in real time with a monocular camera, which is relevant to search and rescue operations. The paper's strengths are the integration of modern CNNs in a ROS2 pipeline, the use of data-driven system identification for controller tuning, and the public availability of the implementation and test videos. However, the central quantitative claim of maintaining a specified 2 m distance is not currently supported: the distance estimator is not validated against ground truth, the controller gains and identified model parameters are not reported, and the experimental section contains no quantitative accuracy metrics for detection, recognition, or tracking.","major_comments":[{"comment":"The claim that the UAV maintains a specified distance (2 m, Section 3.5) is load-bearing and currently unsupported. Equation (1) is the sole distance signal for all three PD controllers, yet the paper reports no fitted coefficients k1, k2, and k3, no calibration protocol, and no comparison against measured distances. Add the calibration procedure, the coefficient values, and error statistics (e.g., RMSE or MAE against a measured range over 0–6 m) to substantiate the metric-distance claim.","section":"Section 4, Eq. (1)"},{"comment":"The experimental evaluation is entirely qualitative: there are no precision/recall or mAP numbers for people detection, no recognition accuracy or false-accept/reject rates for face recognition, and no tracking error or distance-hold statistics. The statement that 'all the proposed deep learning methods demonstrated excellent performance' is not measurable. Report quantitative metrics with defined success criteria for detection, recognition, and tracking.","section":"Section 4"},{"comment":"The number of test individuals is inconsistent: the abstract states that experiments were conducted on 14 known individuals, while Section 4 states that the system was tested on 11 individuals. Clarify which number is correct and report the number of trials, the environments used, and any exclusion criteria.","section":"Abstract vs. Section 4"},{"comment":"The system identification and PD tuning are described only in words. The identified transfer functions or state-space models, the PD gains, and the Simulink closed-loop validation (nominally shown in Fig. 2) are not provided. Without these values, the 'data-driven system identification' contribution cannot be assessed. Report the identified model parameters and gains, along with a model-fit metric (e.g., fit percentage) and a closed-loop performance measure.","section":"Section 3.1"},{"comment":"Equation (1) assumes a stable shoulder-to-hip reference length and average European heights (180 cm male, 171 cm female). For individuals of different height or body proportions, and for postures such as leaning, the distance estimate is systematically biased, and the controllers treat this biased estimate as truth. The paper acknowledges leaning as a limitation but does not quantify its effect or validate the estimator across different subjects and poses. This is load-bearing for the 'maintains a specified distance' claim and should be addressed with multi-subject, ground-truth validation.","section":"Section 3.4, Eq. (1)"}],"minor_comments":[{"comment":"The paper uses dlib's face recognition model but cites reference [11], which describes FaceNet; dlib's model is a different ResNet-based architecture, so the citation should be to dlib's documentation or to an appropriate source for that model.","section":"Section 3.3"},{"comment":"The processing rates are inconsistent: Section 3.2 states real-time detection at 15 Hz, while Section 4 states that YOLOv11 processes frames at 30 Hz and YOLOv11-pose at 15 Hz. Reconcile these numbers.","section":"Sections 3.2 and 4"},{"comment":"The graph labels 'graph1' and 'graph2' are not descriptive; relabel the subplots as input signal and output velocity for clarity.","section":"Fig. 2"},{"comment":"The terminology for the controlled axes is confusing: the text refers to a 'roll/X-axis PD controller' and later to 'yaw/Z-axis linear movement.' Use consistent notation for translational axes (X, Z) and rotational axes (yaw).","section":"Section 3.1"},{"comment":"The phrase 'losing traction and failing to maintain tracking procedure' appears to be a typo for 'losing track'; please correct the wording.","section":"Section 4"},{"comment":"Several references are incomplete or inconsistent with a journal style, such as [5] lacking page numbers and [10] lacking volume and page details; please check all entries against the target journal's reference format.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is essentially a system integration report, and its value depends on the quantitative validation the authors can add. The inconsistency in participant count and the absence of metrics suggest the manuscript was submitted prematurely. I would encourage the editor to require a proper experimental evaluation with ground-truth distance validation before acceptance; the integration itself appears plausible and the public release of code and videos is a point in its favor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real, working proof of concept, and the stress-test concern is correct. The authors built a ROS2 pipeline on a DJI Tello with a Jetson Xavier NX: YOLOv11 detects people, dlib recognizes faces, YOLOv11-pose extracts body keypoints, and three PD controllers keep the chosen individual centered and at a set distance. Getting that loop to run in real time on a small UAV is an actual engineering achievement, and the GitHub repo with videos is a plus.\n\nWhat is new is the use of shoulder-to-hip keypoints, rather than the face, as the distance reference, with a fitted quadratic mapping from pixel span to centimeters. That is a modest extension of prior face-tracking work, but it is genuinely useful for scenarios where the face is not visible, and the system-level integration of YOLOv11 + dlib + YOLOv11-pose is not in the cited literature.\n\nThe soft spots are real. Equation (1) is the single distance signal feeding all three controllers. The authors give no fitted k1, k2, k3, no calibration protocol, and no ground-truth comparison between estimated and measured distance. The claim that the drone maintains two meters is therefore not substantiated. The calibration assumes an average European height, which will bias the estimate for shorter and taller people, and the authors themselves note that leaning distorts the shoulder-to-hip span. None of this is fatal for a proof of concept, but it means the central quantitative claim is unvalidated.\n\nTwo smaller issues: the abstract says 14 known individuals while Section 4 says 11, and the paper reports no precision/recall or error metrics for detection, recognition, or tracking. For an engineering evaluation, that is the main gap.\n\nWho benefits: anyone working on UAV person-following or search-and-rescue systems will get a clear picture of a functioning baseline. The paper deserves peer review rather than desk rejection, because the system is real, the code appears available, and the missing pieces are fixable. My recommendation: send it out—workshop or applied venue—with a clear request for major revisions: report the fitted coefficients, validate Eq. (1) against ground truth or known-height subjects, and add quantitative tracking errors. I would not cite the 2 m tracking claim until that is done, but I would cite the paper as related work on pose-based distance control.","headline":"A credible real-time UAV detect-recognize-track pipeline, but the load-bearing distance estimator is never validated against ground truth, so the 2 m tracking claim is unsupported.","tokens_in":6784,"tokens_out":3885,"would_cite":true,"duration_ms":40277,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a small drone can search for a known person, recognize their face, and follow them at a safe distance using only a single camera and neural networks.","keywords":["UAV","search and rescue","computer vision","convolutional neural network","YOLOv11","face recognition","pose estimation","PD control"],"falsifier":"Place people of known different heights, say 160 cm and 190 cm, at measured distances from the UAV, hold posture fixed, and compare the system's reported distance with a tape measure or a distance sensor; if the error grows with height or with a 20-degree lean while true distance is constant, the distance signal is biased.","tokens_in":5855,"feed_emoji":"🚁","tokens_out":10619,"duration_ms":101997,"temperature":0.7,"pith_summary":"This paper is trying to show that a small quadcopter can perform a complete search-and-rescue subtask on its own: detect a person, recognize whether that person is the one being sought, and then follow them while keeping a set distance. The system runs on a DJI Tello UAV with video streamed to a Jetson Xavier NX board, using YOLOv11 to spot people, a dlib face-recognition network to identify the target, and YOLOv11-pose to estimate body keypoints. Distance to the person is inferred from the pixel length of the shoulder-to-hip segment through a calibrated quadratic formula, and three PD controllers use that distance to drive forward, vertical, and yaw motion. If correct, this provides a compact proof of concept for autonomous UAV assistance in search and rescue, with the perception happening off the ground station.","feed_headline":"Small drone detects, recognizes, and follows a target person","feed_subtitle":"Body-keypoint distance lets a Tello quadcopter hold a standoff while tracking a recognized person.","key_machinery":"The load-bearing mechanism is the shoulder-to-hip keypoint segment extracted by YOLOv11-pose. Its pixel length $x$ is mapped to a physical distance estimate $y = k_1 x^2 + k_2 x + k_3$ in centimeters, calibrated experimentally from 0 to 6 meters on average European body heights. Because shoulders and hips remain visible from many orientations, this single monocular cue lets the UAV track a person whose face is not visible, and it supplies the error signal to three PD controllers governing X translation, Z translation, and yaw. The system identification step provides the linear dynamic models for each axis that make the PD gains appropriate.","core_discovery":"The central discovery is that the full detect-recognize-track loop can be closed on a small UAV using only a monocular camera. YOLOv11 filters its detections to people, dlib's face embedding compares each detected face against a stored template, and YOLOv11-pose returns 17 body keypoints. The shoulder-to-hip pixel distance is converted to physical distance with a fitted quadratic, and that distance plus the shoulder midpoint's offset from the image center feed three PD controllers for forward/backward, up/down, and yaw. The PD gains come from data-driven system identification of the Tello's dynamics using IMU telemetry. In tests the loop ran in real time, with face recognition only in the search phase and tracking following the body rather than the face.","pith_inferences":["The distance estimate has not been checked against ground truth; a direct range test with people of different heights and body angles would show whether the controllers are holding physical distance or merely keeping a pixel size constant.","Because the shoulder-to-hip calibration is tied to average European body dimensions, the same system would need re-calibration for populations with different height distributions.","The pose model runs at half the frame rate of detection, so the limiting factor for tracking smoothness is likely the pose/control chain rather than the person detector; raising pose estimation to the same rate as detection is a testable improvement."],"forward_implications":["The UAV can follow a person from angles where the face is hidden, as long as shoulders and hips are visible.","An operator can initialize the system without a stored photo by centering the person and saving a template mid-flight.","The loop's timing budget is set by control at 15 Hz and pose at 15 Hz, which is adequate for slow human motion.","If a Wi-Fi glitch interrupts tracking, the system can re-run face recognition against stored templates to reacquire the target.","The same architecture can be ported to a larger UAV with better cameras and global positioning, as the authors state as next steps."],"supporting_citations":[{"why":"Provides the DJI Tello face-tracking PID baseline that this system replaces with body-keypoint distance and tuned PD controllers.","marker":"[6]"},{"why":"Introduces the keypoint-plus-PD-control tracking architecture that the paper extends from facial keypoints to shoulder and hip keypoints.","marker":"[9]"},{"why":"Supplies the face-recognition embedding whose 128-dimensional vector and 0.6 threshold are used to identify the searched individual.","marker":"[11]"},{"why":"Documents the DJI Tello platform whose 720p camera, IMU telemetry, and ROS compatibility the whole system builds on.","marker":"[12]"},{"why":"Supplies the pretrained YOLOv11 and YOLOv11-pose models used for person detection and body keypoint extraction.","marker":"[13]"}],"fun_headline_variants":["Drone locks on, identifies, and tracks a person","Tello quadcopter finds a face, then follows the body","Autonomous search drone: detect, recognize, track","Monocular UAV tracks recognized people via body keypoints"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the pixel length of a person's shoulder-to-hip segment is a stable ruler for physical distance, but the quadratic mapping is calibrated on average European heights and does not account for individual height or posture, so a shorter, taller, or leaning person produces a biased distance that the PD controllers treat as true.","fun_headline_variants_meta":{"raw":{"variants":["Drone locks on, identifies, and tracks a person","Tello quadcopter finds a face, then follows the body","Autonomous search drone: detect, recognize, track","Monocular UAV tracks recognized people via body keypoints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000323,"raw_usage":{"total_tokens":1829,"prompt_tokens":972,"completion_tokens":857,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":790}},"tokens_in":588,"tokens_out":857,"duration_ms":9576,"temperature":1.0,"reasoning_tokens":790,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:16:23.432262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Place people of known different heights, say 160 cm and 190 cm, at measured distances from the UAV, hold posture fixed, and compare the system's reported distance with a tape measure or a distance sensor; if the error grows with height or with a 20-degree lean while true distance is constant, the distance signal is biased.","supporting_citations":[{"cited_title":"Face Tracking for Flying Robot Quadcopter based on Haar Cascade Classifier and PID Controller,","cited_arxiv_id":null,"evidence_quote":"Provides the DJI Tello face-tracking PID baseline that this system replaces with body-keypoint distance and tuned PD controllers."},{"cited_title":"Person Tracking and Frontal Face Capture with UAV,","cited_arxiv_id":null,"evidence_quote":"Introduces the keypoint-plus-PD-control tracking architecture that the paper extends from facial keypoints to shoulder and hip keypoints."},{"cited_title":"FaceNet: A Unified Embedding for Face Recognition and Clustering,","cited_arxiv_id":null,"evidence_quote":"Supplies the face-recognition embedding whose 128-dimensional vector and 0.6 threshold are used to identify the searched individual."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the DJI Tello platform whose 720p camera, IMU telemetry, and ROS compatibility the whole system builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained YOLOv11 and YOLOv11-pose models used for person detection and body keypoint extraction."}],"review_version":1}