Pith. sign in

REVIEW 4 major objections 6 minor 39 references

InDT replaces CFAR thresholding with a deep detector over three-channel Range-Doppler inputs, reporting about a 10 dB detection gain over CA-CFAR and lower OSPA tracking error than PMBM.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 23:29 UTC pith:X2MQMDY7

load-bearing objection A plausible learned-detection-plus-KF pipeline for radar, but the headline 10 dB gain rests on an unspecified Pfa-matching basis and the tracking evaluation is too thin to carry the OSPA claim. the 4 major comments →

arxiv 2509.06569 v1 pith:X2MQMDY7 submitted 2025-09-08 eess.SP cs.AI

Integrated Detection and Tracking Based on Radar Range-Doppler Feature

Chenyu Zhang , Yuanhang Wu , Xiaoxi Ma , Wei Yi This is my paper
classification eess.SP cs.AI
keywords radar detectionRange-Dopplerintegrated detection and trackingdeep learning detectoradaptive Kalman filterdata associationCFARmulti-target tracking
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper proposes InDT, a learning-based replacement for the classical radar chain of constant-false-alarm-rate detection followed by Bayesian filtering. The detector consumes a three-channel Range-Doppler representation — normalized amplitude, real part, and imaginary part — so phase information survives thresholding, and a feature-enhancement network outputs target positions and confidence scores. The tracker uses the confidence scores to scale the Kalman filter's measurement noise covariance and combines cosine distance between RD features with Mahalanobis distance for data association. The paper reports about 10 dB and 4 dB detection gains over CA-CFAR and a Monte Carlo threshold method at the same approximate false-alarm rate, and lower OSPA error than PMBM in low-SNR multi-target and measured automotive scenarios. If those gains hold under matched false-alarm conditions, InDT is a practical way to detect and track weak radar targets without manual threshold tuning.

Core claim

The central claim is that an integrated, learned detector can exploit signal structure — especially phase — that threshold detectors discard, and that the extra information improves both detection and tracking. InDT maps each RD matrix to dense features with convolutional layers, enhances them with self-attention over local windows, and regresses target positions plus a confidence score; the tracker turns the confidence into an adaptive measurement-noise covariance for a Kalman filter and combines location with stored feature vectors via cosine distance during association. On simulated data at about -20 dB SNR the paper obtains detection gains of about 10 dB over CA-CFAR and 4 dB over a Mont

What carries the argument

The load-bearing mechanism is the three-channel Range-Doppler input paired with a confidence loop from detector to tracker. Feeding normalized amplitude, real, and imaginary matrices preserves the complex signal phase that CFAR loses; the trained encoder, built from convolutional layers and local-window self-attention, produces both target positions and a confidence score. That confidence drives a reciprocal scaling of the Kalman measurement-noise covariance, and feature vectors, updated by exponential moving average, contribute a cosine-distance term to the Hungarian assignment cost. The reported gains come from these two information transfers rather than from any single network block.

Load-bearing premise

The false-alarm rates for InDT and the baselines are assumed to be measured on the same basis, but the paper never describes how InDT's confidence threshold is chosen to match Pfa; if the rate is defined differently or the threshold is tuned on the test set, the claimed 10 dB and 4 dB detection gains would be inflated.

What would settle it

Rerun the authors' SNR sweep with the InDT confidence threshold fixed on a validation split, then compute per-cell false-alarm rate from InDT and CA-CFAR outputs on the same held-out RD matrices; if InDT's detection probability at the matched Pfa no longer shows roughly 10 dB better SNR performance, the central detection claim is false.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Radar processing chains can replace the CFAR threshold with one network that outputs positions, confidence, and features for tracking.
  • Weak targets near -20 dB pre-processing SNR become detectable while keeping false alarms at the same level, extending range or cutting transmit power for automotive radar.
  • Detection confidence can serve as a proxy for measurement noise in a Kalman filter, improving track accuracy without redesigning the motion model.
  • RD feature similarity supports data association where location alone fails, particularly under heavy clutter with thousands of false alarms per frame.
  • The same detector-tracker pipeline transfers to measured 77 GHz automotive data and requires no per-scene re-tuning of thresholds.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The headline 10 dB gain is conditional on matched false-alarm definitions; a strict per-cell Pfa comparison with the confidence threshold fixed on a validation set would tell whether the gain survives.
  • The three-channel input implies phase information is a main source of improvement over single-amplitude CFAR; an ablation that removes the imaginary channel would quantify that contribution directly.
  • The reciprocal confidence-to-covariance map is a placeholder; learning the map from detector statistics or using a Bayesian uncertainty estimate could push tracking accuracy beyond the paper's reported numbers.
  • The same integration pattern could be applied to other radar grids, such as MIMO or SAR images, wherever CFAR-style thresholding discards phase and structure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes InDT, an integrated detection-and-tracking pipeline for radar Range-Doppler (RD) matrices. A neural detector with a three-channel real/imaginary/amplitude input extracts and enhances RD features and outputs target positions and confidences; the tracker uses detection confidence to adapt the measurement-noise covariance of a Kalman filter (C-AKF) and augments Mahalanobis-distance association with cosine similarity of learned RD features. Experiments on simulated data report detection gains of about 10 dB over CA-CFAR and 4 dB over a Monte Carlo threshold method at an allegedly equal false-alarm rate, and lower OSPA than PMBM in a low-SNR multi-target scenario. A public automotive dataset is used for detection-rate and two-target tracking comparisons.

Significance. If the detection gains hold under a properly matched false-alarm definition, InDT would be a practically relevant alternative to CFAR-based detection plus Bayesian filtering in low-SNR automotive radar. The three-channel complex/real/amplitude input and the confidence-adaptive Kalman filter are interesting design choices, and evaluating on publicly available raw radar data is a strength. However, the central comparisons are not yet supported: the Pfa-matching protocol is unspecified, the real-data ground truth is generated by the same CFAR algorithm used as baseline, and the tracking evidence rests on a single short simulation and a two-target real sequence without ablations or statistical uncertainty.

major comments (4)
  1. [Sec. 4.1, Fig. 5; Sec. 4.2] The headline detection gain ('about 10 dB and 4 dB over CA-CFAR and Monte Carlo 2') rests on a false-alarm-rate match that is never defined. The abstract calls the Pfa 'approximate'; Sec. 4.2 says only that the confidence threshold is 'adjusted' to match. For CA-CFAR/Monte Carlo, Pfa is naturally a per-range-Doppler-cell threshold-crossing rate; for a neural detector that outputs detections after peak selection/NMS, a false alarm can instead mean a spurious object, and the two definitions can differ by orders of magnitude. Please specify (i) the unit over which Pfa is computed (cell, connected component, or object), (ii) how the InDT confidence threshold is swept and selected (ideally on validation data, not the test set), and (iii) whether the clutter model and measurement density are identical across methods. Without this, the 10/4 dB claim and the subsequent PMBM tracking comparison a
  2. [Sec. 5, Table 3] The real-data ground truth is generated by CFAR detections with manual calibration, and the same CFAR is then used as the baseline in Table 3. This is partially circular: any target missed by CFAR tends to be absent from the ground-truth label, so CFAR's detection rate is inflated and InDT is evaluated only on CFAR-discovered objects. Please use independent annotations (for instance, the synchronized camera labels) or at least report the number/manual corrections applied and provide detection ROC curves against an independent ground-truth set.
  3. [Sec. 4.2, Fig. 6(d); Sec. 5, Fig. 8(b)] The tracking evaluation is a single 30-frame synthetic scenario with no error bars and a single 31-frame, two-target real sequence. The statement that the 'overall OSPA error for InDT is lower' needs a defined aggregate (mean over time? total sum?) and a measure of variability across repeated trials. There is also no ablation study: removing C-AKF or the cosine-feature association would establish that these components, rather than the detector's favorable operating point, cause the tracking improvement.
  4. [Sec. 3.3, Eq. (13)] The binary cross-entropy term is written with the model prediction o_i^t as the coefficient of log(c_i^t); as defined in the text, o_i^t is the predicted probability that the sample is positive. Standard BCE should use the ground-truth indicator as the coefficient. Please correct the notation or the definition, and restate the loss so that the supervised training is unambiguous.
minor comments (6)
  1. [Sec. 4.1] Monte Carlo 1 and Monte Carlo 2 are not fully specified. State the fixed false-alarm values, the clutter model, and how the threshold is computed for each; also clarify the relation between the reported Pfa and the threshold-selection procedure.
  2. [Sec. 3.4, Eq. (15)] If n=0 (no detections in a frame), the adaptive update is undefined. Clarify how empty frames are handled in C-AKF.
  3. [Sec. 4.1, Fig. 4] The text says 'About 7dB after signal processing' for SNR=-20 dB. State explicitly that this is after the 27.09 dB pulse-compression/coherent-integration gain and define the SNR axis.
  4. [Sec. 3.2, Eq. (7)] The notation T(F^t+P, F^t+P) is ambiguous for self-attention. Use standard Q/K/V notation or explain that the same input is used for query, key, and value.
  5. [Sec. 3.3, Eq. (12)] Specify whether the position loss is an MSE over range and Doppler and how the ground-truth state vector y_i^t is encoded.
  6. [Throughout] Typographical and minor language issues: 'Transfomer' (Sec. 3.2), 'Mahalanoulian' (Sec. 3.5), 'trace result vector' (Sec. 3), 'vecotor' (Sec. 3.3). Please proofread.

Circularity Check

1 steps flagged

Real-data detection comparison is partly circular: CFAR-generated ground truth is used to evaluate InDT against CFAR; simulated headline results remain independent.

specific steps
  1. self definitional [Section 5, 'Experimental Results', first paragraph (ground-truth generation for Table 3)]
    "We use the CFAR to maintain a high detection rate to generate radar RD detections, which are manually calibrated and saved as the ground truth required for training and evaluation of the detection task."

    Table 3 compares InDT and CFAR detection rates using ground truth formed from CFAR detections. Because the baseline CFAR produces the candidate detections from which the labels are drawn, any target missed by CFAR cannot appear in the ground-truth set. InDT is therefore trained and scored on CFAR's own detection decisions (plus manual edits), so the reported real-data improvement over CFAR (e.g., 0.995 vs 0.999, 0.881 vs 0.898) is partly by construction. Manual calibration does not remove the dependence, because it cannot recover detections CFAR never made. This circularity is load-bearing only for the Section 5 measured-data validation; the Section 4 simulated detection and tracking experiments are generated independently.

full rationale

Most of the paper's derivation chain is empirical rather than derivational. The detector is trained with supervised losses on simulated data with known ground-truth positions, and the simulated detection gain over CA-CFAR and Monte Carlo thresholds (Section 4.1) is evaluated on independently generated test matrices; the 10 dB and 4 dB claims are not constructed from the baselines. The tracking comparison with PMBM uses independently defined trajectories and OSPA, and matching InDT's false-alarm scenario to PMBM's is an operating-point choice rather than a fitted-input-called-prediction. No load-bearing self-citation chain or imported uniqueness theorem is present. The one genuine circular element is the measured-data detection ground truth in Section 5: CFAR outputs are manually calibrated into the labels against which CFAR is later compared in Table 3. Since the label set inherits the baseline's detections, the real-data 'outperforms CFAR' claim is partially self-referential. The paper also leaves the confidence-to-Pfa mapping unspecified ('approximate false alarm rate' in the abstract, 'adjust the detection confidence of InDT to match the same false alarm scenario' in Section 4.2), which is a measurement-protocol weakness and a potential fairness concern, but not a demonstrated circular reduction. Overall, the central simulated results are independent, so the circularity is bounded and moderate.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central results rest on a set of hand-chosen hyperparameters and an ad hoc covariance adaptation rule; no new physical entities are introduced, and the simulation and filter models assume standard Gaussian noise and constant-velocity motion.

free parameters (6)
  • lambda_1 = 0.7
    Position loss weight in Eq. (11), set by hand.
  • lambda_2 = 0.3
    Confidence loss weight in Eq. (11), set by hand.
  • alpha = 0.7
    EMA feature update factor in Eq. (18), set by hand.
  • mu = 0.3
    Weight for combining Mahalanobis and cosine distances in Eq. (19), set by hand.
  • InDT confidence threshold = not reported
    Threshold matching Pfa to CFAR/Monte Carlo baselines; selection procedure not described in Section 4.1.
  • C-AKF scaling function = n^2 / sum(c_i)
    Ad hoc mapping in Eq. (15) from confidence to measurement noise covariance; no derivation, acknowledged as open question.
axioms (4)
  • domain assumption Radar echo model with Gaussian white noise and constant-velocity targets (Eqs. 1-3)
    Used to generate simulation data and to justify the CFAR baseline; real clutter may not be Gaussian.
  • domain assumption Constant velocity motion model for targets (Eq. 20)
    Used for the Kalman filter; real targets may maneuver, which would degrade tracking.
  • ad hoc to paper Detection confidence is inversely related to measurement noise covariance (Section 3.4)
    The C-AKF update rule is introduced without principled justification; the paper states finding a better function is open.
  • domain assumption Public dataset ground truth from CFAR plus manual labeling is accurate (Section 5)
    If CFAR missed targets or the manual calibration is biased, the reported detection rates for InDT on real data are unreliable.

pith-pipeline@v1.3.0-alltime-deepseek · 9653 in / 17587 out tokens · 173857 ms · 2026-08-04T23:29:00.143709+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Integrated Detection and Tracking Based on Radar Range-Doppler Feature." pith.science (2026). https://pith.science/paper/X2MQMDY7

@misc{pith2026250906569,
  author       = {Pith},
  title        = {Pith review of: Integrated Detection and Tracking Based on Radar Range-Doppler Feature},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X2MQMDY7}},
  note         = {Machine review of arXiv:2509.06569}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Detection and tracking are the basic tasks of radar systems. Current joint detection tracking methods, which focus on dynamically adjusting detection thresholds from tracking results, still present challenges in fully utilizing the potential of radar signals. These are mainly reflected in the limited capacity of the constant false-alarm rate model to accurately represent information, the insufficient depiction of complex scenes, and the limited information acquired by the tracker. We introduce the Integrated Detection and Tracking based on radar feature (InDT) method, which comprises a network architecture for radar signal detection and a tracker that leverages detection assistance. The InDT detector extracts feature information from each Range-Doppler (RD) matrix and then returns the target position through the feature enhancement module and the detection head. The InDT tracker adaptively updates the measurement noise covariance of the Kalman filter based on detection confidence. The similarity of target RD features is measured by cosine distance, which enhances the data association process by combining location and feature information. Finally, the efficacy of the proposed method was validated through testing on both simulated data and publicly available datasets.

Figures

Figures reproduced from arXiv: 2509.06569 by Chenyu Zhang, Wei Yi, Xiaoxi Ma, Yuanhang Wu.

Figure 1
Figure 1. Figure 1: Conceptual comparison of the integrated detection and tracking methods. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: InDT consists of 4 main components: (1) A feature encoder that extracts the feature vector of each signal [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A model input method for three-channel complex value signal splicing. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of detection: (a) SNR=−20dB signal matrix detection result. (b) Amplitude channel feature map. (c) Real-valued channel feature map. (d) Complex-valued channel feature. Data generation. The simulation data follows the radar processing procedure in Section 2 to generate the raw radar echo data with different SNR. We use the Linear Frequency Modulated Continuous Wave (LFMCW) signal and the param… view at source ↗
Figure 5
Figure 5. Figure 5: Detection performance: (a) The Pd are plotted against SNR. (b) The corresponding Pfa are plotted against SNR. The Monte Carlo threshold method approximates the theoretical upper limit of the threshold detection algorithm. In the simulation experiment, we set a fixed false alarm value to limit the number of false alarms that can be detected. All clutter energies are sorted, and the clutter energy at the jun… view at source ↗
Figure 6
Figure 6. Figure 6: Multi-target tracking scenario with low SNR. (a) Single frame measurement. (b) Single frame ground truth. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Columns 1-3 show the car park scene, and the fourth column shows the roadside scene. For each column, the [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: (a) Visualisation of target trajectories with synchronised camera images. (b) The OSPAs of range (m) are [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 31 canonical work pages

  1. [1]

    Radar CFAR thresholding in clutter and multiple target situations.IEEE Trans

    Hermann Rohling. Radar CFAR thresholding in clutter and multiple target situations.IEEE Trans. Aerosp. Electron. Syst., (4):608–621, 1983

    work page 1983

  2. [2]

    R. E. Kalman. A new approach to linear filtering and prediction problems.Journal of Basic Engineering, pages 35–45, 1960

    work page 1960

  3. [3]

    Performance evaluation of radar detection schemes based on CA-CFAR against K-distributed clutter

    Pierfrancesco Sciotti, Massimo;Lombardo. Performance evaluation of radar detection schemes based on CA-CFAR against K-distributed clutter. In2001 CIE International Conference on Radar Proceedings, 2001

    work page 2001

  4. [4]

    probability hypothesis density (PHD)

    Ronald P Mahler. A theoretical foundation for the stein-winter" probability hypothesis density (PHD)" multitarget tracking approach. Technical report, Army Research Office Alexandria Va, 2000

    work page 2000

  5. [5]

    García-Fernández, Jason L

    Ángel F. García-Fernández, Jason L. Williams, Karl Granström, and Lennart Svensson. Poisson multi-bernoulli mixture filter: Direct derivation and implementation.IEEE Trans. Aerosp. Electron. Syst., 54(4):1883–1901, 2018

    work page 1901

  6. [6]

    Integration of bayes detection with target tracking.IEEE Trans

    Peter Willett, Ruixin Niu, and Yaakov Bar-Shalom. Integration of bayes detection with target tracking.IEEE Trans. Signal Process., 49(1):17–29, 2001

    work page 2001

  7. [7]

    Offline performance prediction of pdaf with bayesian detection for tracking in clutter.IEEE Trans

    Tao Zeng, Le Zheng, Yang Li, Xinliang Chen, and Teng Long. Offline performance prediction of pdaf with bayesian detection for tracking in clutter.IEEE Trans. Signal Process., 61(3):770–781, 2012

    work page 2012

  8. [8]

    Joint detection and tracking processing algorithm for target tracking in multiple radar system.IEEE Sensors Journal, 15(11):6534–6541, 2015

    Junkun Yan, Hongwei Liu, Bo Jiu, Zheng Liu, and Zheng Bao. Joint detection and tracking processing algorithm for target tracking in multiple radar system.IEEE Sensors Journal, 15(11):6534–6541, 2015

    work page 2015

  9. [9]

    Joint detection and tracking scheme for target tracking in moving platform

    Yongsheng Guan and Yingping Wang. Joint detection and tracking scheme for target tracking in moving platform. InProc. IEEE Radar Conf., pages 1–4, 2020

    work page 2020

  10. [10]

    Signal structure information-based target detection with a fully convolutional network.Information Sciences, 576:345–354, 2021

    Chang Gao, Junkun Yan, Xiaojun Peng, and Hongwei Liu. Signal structure information-based target detection with a fully convolutional network.Information Sciences, 576:345–354, 2021

    work page 2021

  11. [11]

    Data association for maneuvering targets through a combined siamese network and xgboost model.Signal Processing, 211:109086, 2023

    Chang Gao, Junkun Yan, Bo Chen, Pramod K Varshney, Tianyi Jia, and Hongwei Liu. Data association for maneuvering targets through a combined siamese network and xgboost model.Signal Processing, 211:109086, 2023

    work page 2023

  12. [12]

    Tonissen and R.J

    S.M. Tonissen and R.J. Evans. Peformance of dynamic programming techniques for track-before-detect.IEEE Trans. Aerosp. Electron. Syst., 32(4):1440–1451, 1996. 10 Running Title for Header

    work page 1996

  13. [13]

    Recursive bayesian filtering for multitarget track-before-detect in passive radars.IEEE Trans

    Frederic Lehmann. Recursive bayesian filtering for multitarget track-before-detect in passive radars.IEEE Trans. Aerosp. Electron. Syst., 48(3):2458–2480, 2012

    work page 2012

  14. [14]

    Supervised learning based online filters for targets tracking using radar measurements

    Jie Deng, Wei Yi, Kai Zeng, Qiyun Peng, and Xiaobo Yang. Supervised learning based online filters for targets tracking using radar measurements. InProc. IEEE Radar Conf., pages 1–6, 2020

    work page 2020

  15. [15]

    A fast and robust maneuvering target tracking method without Markov assumption

    Chenyu Zhang, Jie Deng, Wei Yi, and Xiujuan Lu. A fast and robust maneuvering target tracking method without Markov assumption. InProc. 25th Int. Conf. Inf. Fusion, pages 1–8, 2022

    work page 2022

  16. [16]

    A study on radar target detection based on deep neural networks.IEEE Sensors Letters, page 7000504, 2019

    Li Wang;Jun Tang;Qingmin Liao. A study on radar target detection based on deep neural networks.IEEE Sensors Letters, page 7000504, 2019

    work page 2019

  17. [17]

    Deep radar detector

    Daniel Brodeski;Igal Bilik;Raja Giryes. Deep radar detector. InProc. IEEE Radar Conf., 2019

    work page 2019

  18. [18]

    Wheeler;Martin Holder;Hermann Winner;Mykel J

    Tim A. Wheeler;Martin Holder;Hermann Winner;Mykel J. Kochenderfer. Deep stochastic radar models. In2017 IEEE Intelligent V ehicles Symposium (IV), 2017

    work page 2017

  19. [19]

    Next generation multitarget trackers: Random finite set methods vs transformer-based deep learning

    Juliano Pinto, Georg Hess, William Ljungbergh, Yuxuan Xia, Lennart Svensson, and Henk Wymeersch. Next generation multitarget trackers: Random finite set methods vs transformer-based deep learning. InProc. 24th Int. Conf. Inf. Fusion, 2021

    work page 2021

  20. [20]

    Simple online and realtime tracking

    Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In IEEE International Conference on Image Processing(ICIP), 2016

    work page 2016

  21. [21]

    Simple online and realtime tracking with a deep association metric

    Dietrich Paulus Nicolai Wojke, Alex Bewley. Simple online and realtime tracking with a deep association metric. InIEEE International Conference on Image Processing (ICIP), 2017

    work page 2017

  22. [22]

    Strongsort: Make DeepSORT great again

    Yunhao Du, Yang Song, Bo Yang, and Yanyun Zhao. Strongsort: Make DeepSORT great again. arXiv:2202.13514, 2022

    Pith/arXiv arXiv 2022

  23. [23]

    Sequence-reversing transform-based coherent integration for high-speed target detection.IEEE Trans

    Xiaolong Li, Guolong Cui, Wei Yi, and Lingjiang Kong. Sequence-reversing transform-based coherent integration for high-speed target detection.IEEE Trans. Aerosp. Electron. Syst., 53(3):1573–1580, 2017

    work page 2017

  24. [24]

    Parameterized centroid frequency-chirp rate distribution for lfm signal analysis and mechanisms of constant delay introduction.IEEE Trans

    Jibin Zheng, Hongwei Liu, and Qing Huo Liu. Parameterized centroid frequency-chirp rate distribution for lfm signal analysis and mechanisms of constant delay introduction.IEEE Trans. Signal Process., 65(24):6435–6447, 2017

    work page 2017

  25. [25]

    A fast approach for detection and parameter estimation of maneuvering target with complex motions in coherent radar system.IEEE Trans

    Zhi Sun, Xiaolong Li, Guolong Cui, Wei Yi, and Lingjiang Kong. A fast approach for detection and parameter estimation of maneuvering target with complex motions in coherent radar system.IEEE Trans. V eh. Technol., 70(10):10278–10292, 2021

    work page 2021

  26. [26]

    A new airborne radar target detection approach based on conditional generative adversarial nets

    Junwei Liu, Jinsheng Xie, Yiwen Nie, Yumiao Wang, and Guolong Cui. A new airborne radar target detection approach based on conditional generative adversarial nets. InProc. IEEE 4th International Conference on Electronics Technology (ICET), pages 184–188, 2021

    work page 2021

  27. [27]

    A machine learning approach to clutter suppression for marine surveillance radar

    Zebiao Wu, Jifang Pei, Weibo Huo, Yulin Huang, Yin Zhang, and Haiguang Yang. A machine learning approach to clutter suppression for marine surveillance radar. InProc. IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pages 3137–3140, 2021

    work page 2021

  28. [28]

    Jifang Pei, Yu Yang, Zebiao Wu, Yanjing Ma, Weibo Huo, Yin Zhang, Yulin Huang, and Jianyu Yang. A sea clutter suppression method based on machine learning approach for marine surveillance radar.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15:3120–3130, 2022

    work page 2022

  29. [29]

    Sigmoid-weighted linear units for neural network function approximation in reinforcement learning.Neural Networks, 107:3–11, 2018

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning.Neural Networks, 107:3–11, 2018

    2018

  30. [30]

    Cspnet: A new backbone that can enhance learning capability of cnn

    Chien-Yao Wang, Hong-Yuan Mark Liao, Yueh-Hua Wu, Ping-Yang Chen, Jun-Wei Hsieh, and I-Hau Yeh. Cspnet: A new backbone that can enhance learning capability of cnn. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 390–391, 2020

    2020

  31. [31]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  32. [32]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  33. [33]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. InEuropean conference on computer vision, pages 213–229, 2020

    2020

  34. [34]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012–10022, 2021. 11 Running Title for Header

    work page 2021

  35. [35]

    Layer normalization.arXiv preprint arXiv:1607.06450, 2016

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization.arXiv preprint arXiv:1607.06450, 2016

    Pith/arXiv arXiv 2016

  36. [36]

    Spatial pyramid pooling in deep convolutional networks for visual recognition.IEEE Trans

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition.IEEE Trans. Pattern Anal. Mach. Intell., 37(9):1904–1916, 2015

    work page 1904

  37. [37]

    Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors

    Hong-Yuan Mark Liao Chien-Yao Wang, Alexey Bochkovskiy. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. arXiv:2207.02696, 2022

    Pith/arXiv arXiv 2022

  38. [38]

    Joint detection and estimation of multiple objects from image observations.IEEE Trans

    Ba-Ngu V o, Ba-Tuong V o, Nam-Trung Pham, and David Suter. Joint detection and estimation of multiple objects from image observations.IEEE Trans. Signal Process., 58(10):5129–5141, 2010

    work page 2010

  39. [39]

    Raw adc data of 77ghz mmwave radar for automotive object detection, 2022

    Xiangyu Gao, Youchen Luo, Guanbin Xing, Sumit Roy, and Hui Liu. Raw adc data of 77ghz mmwave radar for automotive object detection, 2022. 12

    work page 2022