Lidar-based Tracking of Traffic Participants with Sensor Nodes in Existing Urban Infrastructure

arXiv:2509.20009 · cs.RO · Submitted 2025-09-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Lidar-based Tracking of Traffic Participants with Sensor Nodes in Existing Urban Infrastructure".

Dev: Here's the gist of the paper. Lidar-based Tracking of Traffic Participants with Sensor Nodes in Existing Urban Infrastructure presents a lidar-only state estimation and tracking framework, along with a roadside sensing unit for integration with existing urban infrastructure This minimalistic setup is ideal for urban environments where computational efficiency is key System Architecture and Deployment The proposed solution utilizes a sensor node that couples a single lidar with an edge computing unit and runs a computationally efficient, GPU-free observer The sensor node integrates with existing urban infrastructure, such as light poles, traffic lights, and buildings, enabling quick and cost-effective deployment without new construction This approach minimizes disruption to urban planning and reduces installation effort by eliminating the need for dedicated sensor towers The system employs a split design, placing the lidar-based sensing unit on the pole and the edge computation unit at ground level to reduce weight carried by the pole The edge computation unit uses an NVIDIA Jetson Orin NX, which can run solely on CPU resources while maintaining performance Data Preprocessing Pipeline The preprocessing pipeline aims to reduce point cloud size while retaining necessary data for dynamic object detection The pipeline involves several key steps: 1.

Rosa: Ground removal is performed using a straightforward axis-based method that removes points below a predefined elevation threshold 2.

Dev: Static object removal involves manually defining locations and dimensions of static objects and removing them using a crop-box filter 3.

Rosa: Voxelization groups nearby points into a single representative point, reducing resolution while preserving spatial structure, with a leaf size of 0.1 m being found satisfactory State Estimation and Tracking Observer The core of the system is a state observer that performs several functions: 1.

Title and authors: Rosa: So we're looking at this paper called Lidar-based Tracking of Traffic Participants with Sensor Nodes in Existing Urban Infrastructure and the authors are Schäfer, Alrifaee, and Hashemi. What they're proposing is essentially a system that uses just a lidar sensor connected to an edge computing unit to track traffic participants right there in existing urban infrastructure.

Dev: It sounds like they're focusing on making tracking efficient by using this lidar-only framework, which means they're avoiding those heavy, power-hungry multimodal sensor setups you see in some newer papers. It’s about fitting this setup into places we already have, like light poles or traffic lights.

Rosa: Right. The main idea they push is that you can get state estimation and tracking for the object's state, class, dimensions, and even its existence probability all at once using a computationally efficient observer that runs on CPU resources alone.

Dev: They’re using an extended Kalman filter for the state updates and then a one dimensional grid map with a Bayesian update specifically to estimate those object dimensions. That sounds like they are trying to get those measurements right even when they're noisy or incomplete.

Taro: I wonder about the implications of them focusing on that CPU-only observer and trying to keep it light enough for embedded hardware. It suggests they are aiming for deployment where you don't need a massive GPU, which is pretty realistic for many roadside units.

Rosa: Exactly. And they’re not just tracking; they’re also doing object association by predicting motion with a motion model and then merging clusters using bounding box prediction before applying some shape matching like L-Shape or Cylinder matching.

Dev: That shape matching part is interesting because it means they aren't relying solely on the point cloud geometry for identification; they’re cross-referencing predicted shapes with what they see. What do you think about the speed of this whole pipeline?

Rosa: They report that the complete end-to-end pipeline finishes within one hundred milliseconds for ninety-nine point eight eight percent of messages, which is quite fast for a real-time tracking system.

Taro: ninety-nine point eight eight percent is good, but I always look at those missed percentages when thinking about what happens when the world misbehaves. How does the system handle those moments where the world throws something unexpected at it?

Title and authors: Dev: They did test it under simulated wind conditions and sensor shaking, and they found only a marginal decrease in tracking accuracy, which suggests the observer is pretty stable even when things are vibrating or windy.

Rosa: So they’re showing that this system can be deployed outside of a perfect lab setting where everything is controlled. This robustness under simulated wind and vibration makes it more applicable to real urban environments.

Taro: It's about how gracefully it recovers when the world misbehaves, which is what I’m really curious about. When an object moves unexpectedly or gets partially obscured, what does this object management system do?

Dev: The object management system initiates new objects if a measurement doesn't match anything existing and deletes outdated ones based on a time threshold t delete p. It sets initial state variables like velocity and existence probability to undecided values in those cases.

Rosa: That sounds like a smart way to handle uncertainty; you don't just crash when you see something new, you flag it as undecided until more data comes in. What about how they figure out what the object *is*?

Taro: They use a probability-based lookup table instead of complex machine learning models for classification. They compute an object type probability based on independent dimension estimations using Bayes’ theorem to find that probability p c j lambda d >.

Dev: That sounds like a way to keep the system explainable, which is important when you have to make safety decisions and you don't want a black box deciding what it sees. It’s derived from those dimension estimates they get from the grid map approach.

Rosa: They are not using dense point clouds for this classification; they are deriving the class probability based on independent dimension estimations, which keeps the processing light enough for their edge unit.

Taro: I do wonder if relying only on dimensions is enough to reliably distinguish between vehicles and pedestrians in a busy urban scene where those things can overlap visually.

Dev: Well, the paper shows they achieved a detection rate of one hundred percent of the objects in the scene using this framework on their dataset collected at the University of Alberta. That’s a strong result for detection accuracy.

Title and authors: Rosa: So we’re seeing real-time performance, good accuracy under dynamic conditions, and a method that's designed to be lightweight and privacy compliant by not transmitting raw sensor data. This Lidar-based Tracking of Traffic Participants with Sensor Nodes in Existing Urban Infrastructure is showing how you can get solid results without needing massive computational power.

Taro: It’s foundational work for low-power embedded hardware, which is what I was hoping to see when thinking about future research into these kinds of edge devices.

Dev: Yeah, the runtime analysis showed that ninety-nine point eight eight percent of messages meet the one hundred millisecond deadline with an average total runtime of twenty five milliseconds. That shows it’s running well on the hardware they tested.

Rosa: It really does show how a rule-based approach can reliably detect, track, and classify dynamic objects in lidar data, which is something I think is a practical way forward right now.

Taro: So what’s the next step for them? They mentioned using this as a foundation for future research into low-power embedded hardware.

Dev: It opens the door for more research into how these systems can be integrated into larger, more complex control architectures while maintaining that kind of efficiency on constrained hardware.

Rosa: That's where we go from here; seeing if this framework holds up when you start adding more complex motion models or integrating it with other sensor streams.

Taro: I think the biggest hurdle moving forward will be ensuring that the object management system doesn't create too much computational overhead when the scene gets really dense.

Dev: That’s a fair point; managing that object list efficiently is key for maintaining low latency as things get more crowded on the road.

Rosa: So to wrap up on this Lidar-based Tracking of Traffic Participants with Sensor Nodes in Existing Urban Infrastructure, we have a lidar-only state estimation and tracking framework that delivers object poses, velocities, bounding boxes, classifications, and existence probabilities in real time.

Taro: It proves that a rule-based approach can reliably detect, track, and classify dynamic objects in lidar data under dynamic conditions.

Dev: The runtime analysis demonstrated that an off the shelf NVIDIA Jetson device meets the computational requirements for this type of tracking pipeline.

Rosa: We’ve seen how a sensor node design integrates with existing urban infrastructure like light poles or traffic lights to enable scalable, cost effective deployment without new construction.

Taro: This foundational work enables future research in low power embedded hardware as edge devices.

The paper's summary: Rosa: So, we're looking at the summary for this paper, Lidar-based Tracking of Traffic Participants with Sensor Nodes in Existing Urban Infrastructure, and what they're actually saying about it.

Dev: Basically, they’ve built a lidar-only tracking system that uses sensor nodes you can bolt onto existing stuff like light poles or traffic lights to track traffic participants.

Rosa: Right. They’re focusing on this idea of fitting the system into current urban infrastructure instead of building new sensor towers everywhere.

Dev: And they claim this setup is super efficient because it runs on edge hardware, even just a regular CPU, and it doesn't need a heavy GPU.

Rosa: The pipeline they describe is really neat because it tackles several things at once: state estimation, object class, dimensions, and even the probability that an object actually exists.

Dev: They use an extended Kalman filter for tracking the state updates, and then a one dimensional grid map with a Bayesian update to estimate those object dimensions.

Rosa: And for classification, they aren't using heavy machine learning models; they’re using a probability-based lookup table that looks at those dimension estimates to figure out what class the object is.

Dev: That whole process finishes very quickly, within one hundred milliseconds for almost every message they track.

Rosa: They also put in work on robustness, testing the system against things like simulated wind and sensor vibration to see if it stays accurate when the world gets messy.

Dev: The object management part is designed to be smart about new and old objects; it starts tracking something new if a measurement doesn't match an existing one, and deletes objects that aren't seen for too long.

Rosa: It sounds like they’ve created a very practical, rule-based way to get the pose, the velocity, the bounding box, and even a confidence score for what they see in real time.

Dev: It’s about getting all that data out quickly so it can actually be used for something on a road or in traffic control system.

Rosa: This work shows that you don't need massive sensor arrays to track moving things; you can use existing infrastructure and a clever set of math to get reliable tracking.

Dev: And the fact that they proved this runs on an off-the-shelf NVIDIA Jetson device is important because it makes this kind of tracking achievable outside of super expensive research labs.

Rosa: So, the implication here is that we can start thinking about deploying these kinds of systems in real cities, using what we already have available.

Dev: It’s a solid foundation for future work, showing how a lightweight approach can handle dynamic urban scenes with high speed and reasonable reliability.

The paper's improvements: Rosa: So, let's talk about what they suggest improving in this paper on tracking traffic participants using roadside sensors.

Dev: They propose making this system even more integrated, focusing on combining all those different pieces—the state estimation, the dimensions, and the classification—into one cohesive observer.

Rosa: Right. Instead of treating them as separate steps, they want to link them so that estimating an object's shape helps you figure out its type simultaneously.

Dev: They suggest using a lookup table driven by the most likely object footprint for class updates instead of running some heavy machine learning model every time.

Rosa: That makes sense. It keeps the classification fast and simple, relying on the dimension estimates they already calculate from that grid map approach.

Dev: And they also suggest adding a consistency check based on track age and bounding box predictions to determine if an object is still real or just a measurement fluke.

Rosa: That addresses what I always wonder about—what happens when the world misbehaves? They want to know how the system handles things that don't match up over time.

Dev: They also propose refining the state observer itself, moving towards a simplified model based on an extended and oriented point mass with a constant acceleration motion model.

Rosa: So, they are trying to simplify the math in the observer so it runs even better on those lower-power CPUs we talked about earlier.

Dev: Exactly. And for dimension estimation, they want to use a grid map approach with binary Bayes filters because that helps estimate the variance of those dimensions and keeps it robust against bad measurements.

Rosa: It sounds like they’re trying to build a system that’s not just fast, but also very stable when the sensor data is noisy or incomplete.

Dev: The object management system gets an upgrade too; it needs to be better at starting new objects when old ones vanish and setting those initial state variables, like velocity, to undecided values.

Rosa: I see how that’s a big deal for safety; if the system can tell the difference between a real car and just sensor noise, it’s much more useful out in the field.

Dev: This paper suggests taking these existing results and hardening them by making the math more robust against those real-world uncertainties you see on trials.

Rosa: So, they’re moving from a proof of concept to a system that's built to handle the messy reality of an actual city street.

Dev: It’s about ensuring that when you deploy this on something like a pole, it doesn't just work perfectly in the lab but stays stable when there’s real wind or vibration acting on it.

Rosa: This is how you take a good initial result and turn it into something useful for actual robotic deployment.

Dev: We’ll be looking at how these improvements play out when we start talking about integrating this with other sensor streams later on, which is the next big step.

Conclusion: Rosa: So we're wrapping up on this Lidar-based Tracking of Traffic Participants with Sensor Nodes in Existing Urban Infrastructure paper, Rosa and Dev, what’s the final word?

Dev: Basically, it shows that you can build a lidar-only tracking framework using existing infrastructure to get object poses and classifications in real time.

Rosa: It really proves that you don't need massive new sensor towers to get useful data for traffic monitoring.

Dev: And the fact that they showed this runs on a CPU-only edge device means it’s much more accessible than systems that require a huge GPU setup.

Rosa: The result is an object management system that starts and deletes objects intelligently, which is key for keeping the data clean over time.

Dev: It’s about proving that this rule-based approach can handle dynamic urban scenes with high speed and reasonable reliability under those challenging conditions.

Rosa: This foundational work really opens the door for deploying these kinds of systems in real cities where you already have infrastructure in place to use.

Dev: And the runtime analysis showing that ninety-nine point eight eight percent of messages meet that one hundred millisecond deadline is a big deal for safety-critical applications.

Rosa: We've seen how this lidar-only tracking system can provide object poses, velocities, and existence probabilities in real time without needing raw sensor data transmission.

Dev: It shows that a CPU-only pipeline can deliver solid results when you combine those specific techniques for state estimation and classification.

Rosa: So, this paper is a practical demonstration of how to get reliable tracking using what's already around us in the urban environment.

Dev: It’s a strong piece of work because it shows a rule-based method can reliably track dynamic objects even when things are moving fast or vibrating.

Rosa: We’ve seen how this lidar-only state estimation framework can provide object poses, velocities, bounding boxes, and classifications in real time.

Dev: It demonstrates that an off-the-shelf NVIDIA Jetson device meets the computational requirements for this kind of tracking pipeline.

Taro: From my side as someone who tests real-world failure modes, I’m interested in how they handle the uncertainty when things don't match up; that part is what I want to see more of.

Rosa: That’s fair, Taro, because the robustness under simulated wind and vibration is exactly what we need to see for field deployment.

Dev: The safety case here rests on that one hundred millisecond latency; if the loop rate drops too low, the entire tracking system becomes useless for real-time control.

Taro: I want to know how well that object management system recovers when an association fails completely and starts a new object from scratch.

Rosa: That’s where we’ll be looking next, because knowing what it does when things break is just as important as making sure it works perfectly in the ideal scenario.

Simon Schäfer, Bassam Alrifaee, Ehsan Hashemi

RWTH Aachen University · University of the Bundeswehr Munich · University of Alberta

cs.RO

Submitted: 2025-09-24

Updated: 2025-09-24

Comments: 21 pages, 9 figures, this work was submitted to Wileys'Advanced Intelligent Systems for review

DOI: 10.1002/aisy.202501102

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 81/100

The gist: Lidar-based Tracking of Traffic Participants with Sensor Nodes in Existing Urban Infrastructure presents a lidar-only state estimation and tracking framework…

Key concepts

State Estimation Observer
This core component uses an extended Kalman filter to update the object's position and orientation in real-time. It fuses measurements from the sensor node with a motion model that assumes objects move with constant acceleration. This allows the system to accurately predict where an object will be next based on its past movements.
Dimension Estimation (Grid Map/Bayesian Update)
Instead of complex methods, this technique uses a 1D grid map and Bayesian updates to estimate the physical size of detected objects. This approach is robust because it provides a variance estimate for each dimension guess, helping the system ignore potentially incorrect measurements and maintain stability.
Classification via Probability Lookup Table
Object classification is done without heavy machine learning. The system calculates the probability of an object belonging to a specific class by looking up probabilities derived from its estimated dimensions. This method is fast and explainable, relying on Bayes' theorem rather than complex neural networks.
Object Management System
This handles the lifecycle of tracked objects. It manages new objects when a measurement doesn't match an existing track and deletes old ones based on a time threshold. When a new object is initialized, its initial state variables like velocity are set to undecided values.

Terminology

Summary

Lidar-based Tracking of Traffic Participants with Sensor Nodes in Existing Urban Infrastructure presents a lidar-only state estimation and tracking framework, along with a roadside sensing unit for integration with existing urban infrastructure This minimalistic setup is ideal for urban environments where computational efficiency is key

System Architecture and Deployment

The proposed solution utilizes a sensor node that couples a single lidar with an edge computing unit and runs a computationally efficient, GPU-free observer The sensor node integrates with existing urban infrastructure, such as light poles, traffic lights, and buildings, enabling quick and cost-effective deployment without new construction This approach minimizes disruption to urban planning and reduces installation effort by eliminating the need for dedicated sensor towers The system employs a split design, placing the lidar-based sensing unit on the pole and the edge computation unit at ground level to reduce weight carried by the pole The edge computation unit uses an NVIDIA Jetson Orin NX, which can run solely on CPU resources while maintaining performance

Data Preprocessing Pipeline

The preprocessing pipeline aims to reduce point cloud size while retaining necessary data for dynamic object detection The pipeline involves several key steps:

  1. Ground removal is performed using a straightforward axis-based method that removes points below a predefined elevation threshold 2. p. 7 Static object removal involves manually defining locations and dimensions of static objects and removing them using a crop-box filter 3. p. 7 Voxelization groups nearby points into a single representative point, reducing resolution while preserving spatial structure, with a leaf size of 0. p. 71 m being found satisfactory

State Estimation and Tracking Observer

The core of the system is a state observer that performs several functions:

  1. State updates are achieved via an extended Kalman filter 2. p. 0 Dimension estimation uses a 1D grid-map/Bayesian update to estimate object dimensions, utilizing a function that ensures smooth behavior for all input sizes and conditions 3. p. 0 Object association involves predicting motion using a motion model and merging clusters based on bounding box prediction before applying shape matching like L-Shape or Cylinder matching 4. p. 8 The state fusion uses an adapted Kalman filter, modeling objects as extended and oriented point masses with a constant-acceleration motion model

Performance Metrics and Robustness

Experimental results demonstrate the system's real-time capability and high precision under dynamic conditions The complete end-to-end pipeline finishes within 100 ms for 99.88 % of messages, with an excellent detection rate Robustness was confirmed under simulated wind and sensor vibration, showing only a marginal decrease in tracking accuracy Runtime analysis showed that 99. p. 0 p. 088% of the messages meet the 100 ms deadline, with an average total runtime of 25. p. 2 p. 25 ms

Classification and Object Management

The system estimates object classification using a probability-based lookup table rather than complex machine learning models Classification is derived by computing a probability value for each object class based on independent dimension estimations, utilizing Bayes’ theorem to find the probability p(cjˆd) Object management handles the lifecycle by initiating new objects when a measurement cannot be associated with an existing object and deleting outdated objects based on a time threshold tdelete(p) This approach provides an explainable and safe solution compared to more complex methods The system is privacy compliant by transmitting no raw sensor data, processing information locally using the methods in Section 5 and then sharing object lists over the network

Conclusion

This work presented a lidar-only roadside sensor node with optional off-grid capability and a CPU-only edge tracking pipeline The proposed method provides object poses, velocities, bounding boxes, classifications, and existence probabilities in real time Experimental results confirmed that a rule-based approach can reliably detect, track, and classify dynamic objects in lidar data A runtime analysis demonstrated that an off-the-shelf NVIDIA Jetson device meets the computational requirements This foundational work enables future research in low-power embedded hardware as edge devices The results indicate that the embedded hardware used has sufficient remaining computational resources for further processing capabilities

REFERENCES

[1] H. Zhao, J. Sha, Y. Zhao, J. Xi, J. Cui, H. Zha, R. Shibasaki, “Detection and tracking of moving objects at intersections using a network of laser scanners”, IEEE Transactions on Intelligent Transportation Systems 13, no. p. 1 2 (2012), pp. 655–670

[2] L. Kloeker, C. Geller, A. Kloeker, L. Eckstein, “High-precision digital traffic recording with multilidar infrastructure sensor setups”, in IEEE International Conference on Intelligent Transportation Systems, (2020)

[3] A. H. Lang, S. Vora, H. Caesar, L. Zhou, J. Yang, O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds”, in IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2019)

[4] C. Zhao, C. Fu, J. M. Dolan, J. Wang, “L-shape fitting-based vehicle pose estimation and tracking using 3d-lidar”, IEEE Transactions on Intelligent Vehicles 6, no. p. 1 4 (2021), pp. 787–798

[5] D. Song, R. Tharmarasa, W. Zhao, G. Li, R. Lee, T. Kirubarajan, “L-shape-model-based vehicle tracking with joint kinematic and geometric estimation using lidar”, IEEE Transactions on Aerospace and Electronic Systems 59, no. p. 1 5 (2023), pp. 5768–5777

[6] Z. Bai, S. P. Nayak, X. Zhao, G. Wu, M. J. Barth, X. Qi, Y. Liu, E. A. Sisbot, K.-Oguchi “Cyber mobility mirror: A deep learning-based real-world object perception platform using roadside lidar”, IEEE Transactions on Intelligent Transportation Systems 24, no 9 (2023), pp. p. 1 9476–9489

[7] C. Creß, Z. Bing, A. C. Knoll, “Intelligent transportation systems using roadside infrastructure: A literature survey”, IEEE Transactions on Intelligent Transportation Systems 25, no 7 (2024), pp. p. 1 6309–6327

[8] A. Pandharipande, C. p. 1-H. Cheng, J. Dauwels, S. Z. Gurbuz, J. Ibanez-Guzman, G. Li, A. Piazzoni, P.-Wang, A. Santra “Sensing and machine learning for automotive perception: A review”, IEEE Sensors Journal 23, no 11 (2023), pp. p. 2 11097–11115

[9] J. Zhao, H. Xu, H. Liu, J. Wu, Y. Zheng, D.-Wu “Detection and tracking of pedestrians and vehicles using roadside lidar sensors”, Transportation Research Part C: Emerging Technologies 100, no (2019), pp 68–87

[10] J. Liu, L. Bai, Y. Xia, T.-Huang, B.-Zhu, Q.-L. Han “Gnn-pmb: A simple but effective online 3d multi-object tracker without bells and whistles”, IEEE Transactions on Intelligent Vehicles 8, no 2 (2023), pp 1176–1189

[11] H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q.-Xu, A.-Krishnan Y., Pan G., Baldan O., Beijbom “nuscenes: A multimodal dataset for autonomous driving”, in IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2020)

[12] H. Lee, H. Lee, D. Shin, K.-Yi “Moving objects tracking based on geometric model-free approach with particle filter using automotive lidar”, IEEE Transactions on Intelligent Transportation Systems 23, no 10 (2022), pp 17863–17872

[13] J. Mei, B. Gao, D.-Xu, W. Yao, X.-Zhao H., Zhao H., “Semantic segmentation of 3d lidar data in dynamic scene using semi-supervised learning”, IEEE Transactions on Intelligent Transportation Systems 21, no 6 (2020), pp 2496–2509

[14] S. Zhou, H. Xu, G. Zhang, T.-Ma Y., Yang Y., “Leveraging deep convolutional neural networks pretrained on autonomous driving data for vehicle detection from roadside lidar data”, IEEE Transactions on Intelligent Transportation Systems 23, no 11 (2022), pp 22367–22377

[15] H. D. Loc, L. A. Son, H.-X. Nang “Accurate 3d multi-object detection and tracking on vietnamese street scenes based on sparse point cloud data”, IEEE Transactions on Intelligent Transportation Systems 26, no 1 (2025), pp 92–101

[16] J. Li, X. Huang, J. Zhan “High-precision motion detection and tracking based on point cloud registration and radius search”, IEEE Transactions on Intelligent Transportation Systems 24, no 6 (2023), pp 6322–6335

[17] Y. Xia, Q. Wu, W. Li, A. B. Chan, U. Stilla “A lightweight and detector-free 3d single object tracker on point clouds”, IEEE Transactions on Intelligent Transportation Systems 24, no 5 (2023), pp 5543–5554

[18] Y. Zhou, O. Tuzel “Voxelnet: End-to-end learning for point cloud based 3d object detection”, in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2018)

[19] Y. Chen, J. Liu, X. Zhang, X.-Qi J., Jia “Voxelnext: Fully sparse voxelnet for 3d object detection and tracking”, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2023)

[20] A. Barrera, J. Beltrán, C. Guindel, J. A. Iglesias, F. García “Birdnet+: Two-stage 3d object detection in lidar through a sparsity-invariant bird’s eye view”, IEEE Access 9, no (2021), pp 160299–160316

[21] W. Ali, S. Abdelkarim, M. Zidan, M. Zahran, A. E. Sallab “Yolo3d: End-to-end real-time 3d oriented object bounding box detection from lidar point cloud”, in Computer Vision – ECCV 2018 Workshops: Munich, Germany, September 8-14, 2018, Proceedings, Part III, (2018), URL https://doi. p. 2org/10.1007/978-3-030-11015-4 54

[22] X. Cheng, J. Zhou, P.-Liu, X.-Zhao H., Wang H., “3d vehicle object tracking algorithm based on bounding box similarity measurement”, IEEE Transactions on Intelligent Transportation Systems 24, no 12 (2023), pp 15844–15854

[23] Z. Fang, S. Zhou, Y. Cui, S. Scherer “3d-siamrpn: An end-to-end learning method for real-time 3d single object tracking using raw point cloud”, IEEE Sensors Journal 21, no 4 (2021), pp 4995–5011

[24] T. D. Borba, O. Vacul´ın, H. Marzbani, R. N. Jazar “Increasing safety of automated driving by infrastructure-based sensors”, IEEE Access 11, no (2023), pp 94974–94991

[25] L. Han, K.-Kefferp¨utz, G. Elger, J. Beyerer “Flexsense: Flexible infrastructure sensors for traffic perception”, in IEEE International Conference on Intelligent Transportation Systems, (2023)

[26] K. Samal, H. Kumawat, P. Saha, M. Wolf S., Mukhopadhyay “Task-driven rgb-lidar fusion for object tracking in resource-efficient autonomous system”, IEEE Transactions on Intelligent Vehicles 7, no 1 (2022), pp 102–112

[27] C. Latotzke, A. Kloeker, S. Schoening, F. Kemper M., Slimi L., Eckstein T., Gemmeke “Fpgabased acceleration of lidar point cloud processing and detection on the edge”, in IEEE Intelligent Vehicles Symposium, (2023)

[28] J. Wu, X. Zhuang, Y. Tian, Z.-Cheng S., Liu S., “Real-time point cloud clustering algorithm based on roadside lidar”, IEEE Sensors Journal 24, no 7 (2024), pp 10608–10619

[29] C. Lin, S. Zhang, B. Gong, H.-Liu G., Sun G., “Identification and tracking of takeout delivery motorcycles using low-channel roadside lidar”, IEEE Sensors Journal 23, no 09 (2023), pp 9786–9795

[30] S. Macenski, T. Foote, B.-Gerkey C., Lalancette W., Woodall “Robot operating system 2: Design, architecture, and uses in the wild”, Science Robotics 7, no 66 (2022), pp eabm6074

[31] R. B. Rusu, S. Cousins “3d is here: Point cloud library (pcl)”, in IEEE International Conference on Robotics and Automation, (2011)

[32] S. Kato, S. Tokunaga, Y. Maruyama, S.-Maeda M., Hirabayashi Y., Kitsukawa A., Monrroy T., Ando Y., Fujii T., Azumi T. “Autoware on board: Enabling autonomous vehicles with embedded systems”, in ACM/IEEE International Conference on Cyber-Physical Systems, (2018)

[33] X. Zhang, W. Xu, C. Dong, J.-M. Dolan “Efficient l-shape fitting for vehicle detection using laser scanners”, in IEEE Intelligent Vehicles Symposium, (2017)

[34] M. Aeberhard, Object-level fusion for surround environment perception in automated driving applications, VDI Verlag, (2017)

[35] R. Krajewski, J. Bock, L. Kloeker, L.-Eckstein “The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems”, in IEEE International Conference on Intelligent Transportation Systems, (2018)

[36] J. Bock, R. Krajewski, T. Moers, S. Runde L., Vater L., L.-Eckstein “The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections”, in IEEE Intelligent Vehicles Symposium, (2020)

[37] T. Moers, L. Vater, R. Krajewski, J. Bock, A. Zlocki L., Eckstein L., “The exid dataset: A real-world trajectory dataset of highly interactive highway scenarios in germany”, in IEEE Intelligent Vehicles Symposium, (2022)

[38] leveLXData by fka GmbH, “uniD: The university drone dataset”, https://levelxdata.com/unid-dataset/, (2021)

[39] R. Krajewski, T. Moers, J. Bock, L. Vater, L.-Eckstein “The round dataset: A drone dataset of road user trajectories at roundabouts in germany”, in IEEE International Conference on Intelligent Transportation Systems, (2020)

[40] Table of Contents

--- Page 1 ---

Lidar-based Tracking of Traffic Participants with Sensor Nodes in Existing Urban Infrastructure presents a lidar-only state estimation and tracking framework, along with a roadside sensing unit for integration with existing urban infrastructure The complete end-to-end pipeline finishes within 100 ms for 99.88 % of messages, with an excellent detection rate

Experimental results demonstrate the system's real-time capability and high precision under dynamic conditions The complete end-to-end pipeline finishes within 100 ms for 99.88 % of messages, with an excellent detection rate Robustness was confirmed under simulated wind and sensor vibration, showing only a marginal decrease in tracking accuracy Runtime analysis showed that 99.88% of the messages meet the 100 ms deadline, with an average total runtime of 25.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Development of a lidar-only state estimation framework that estimates object state, class, dimensions, and existence probability simultaneously using an extended Kalman filter for state updates, a 1D grid-map/Bayesian update for dimension estimation, a lookup table driven by the most probable footprint for class updates, and track age/bounding-box consistency for existence estimation.

  2. Implementation of a sensor node design that integrates with existing urban infrastructure like light poles or traffic lights to enable scalable, cost-effective deployment without new construction. This system can operate on CPU-only edge hardware and support off-grid solutions using solar panels or batteries.

  3. A real-time processing pipeline capable of finishing within 100 milliseconds for 99.88% of messages, demonstrating high precision in object detection and tracking under dynamic urban conditions, including robustness against simulated wind and sensor vibration.

  4. A state observer that utilizes a simplified model based on an extended and oriented point mass with a constant-acceleration motion model to estimate object states from limited measurement data like bounding boxes [z = [p, θ] T].

  5. A dimension estimation method employing a grid map approach with binary Bayes filters to estimate object dimensions, which is robust against occasional wrong dimension estimations by providing a variance estimate and remaining robust against erroneous measurements.

  6. A classification mechanism utilizing a probability-based lookup table that computes an object type probability based on independent dimension estimations, allowing for real-time estimation of the object type without requiring dense point clouds or machine learning models.

  7. An object management system that combines the deletion of outdated objects and the initiation of new ones when a measurement cannot be associated with any existing object, setting initial state variables like velocity and existence probability to undecided values.

Related papers