Connected vehicles and roadside sensors can now "see around" a parked bus by sharing what each of them perceives over a wireless link. This is collaborative perception, and on current benchmarks it adds 10 to 15 points of detection accuracy. Those benchmarks, however, treat every pedestrian as the same box, and the shared data is squeezed hundreds to thousands of times to fit the link. We suspect the gains do not reach the road users who most need them. People in wheelchairs, people using walkers or crutches, and small children have thin, low, unusual shapes. These are the fine details that heavy compression throws away, and they are the shapes a sender's own detector is least likely to recognize and flag. This proposal measures that suspicion and then tests a specific remedy. In the remedy, the receiving vehicle asks for higher-fidelity data only in the places where it is unsure whether a person is present.
The opportunity matters because two literatures each hold half the evidence and neither has asked the joint question. Work on human motion keeps finding that models trained on young, able-bodied people break on prosthesis users, older adults, and mobility-aid users. Work on cooperative driving perception keeps reporting large average gains without ever splitting them by who the pedestrian is. The first deliverable is a stratified benchmark. The second is a test of whether a simple fix closes any gap the benchmark reveals, or whether a more targeted communication rule is needed.
Core Claim
We claim that the cooperative dividend is unevenly distributed. For each pedestrian subgroup $g$ (adult walking, wheelchair, walker, crutches, child-scale) and each per-frame byte budget $b$, define the dividend as
$$D_g(b) = \mathrm{AP}_g(\text{cooperative}, b) - \mathrm{AP}_g(\text{ego only}),$$
where $\mathrm{AP}_g$ is 3D detection average precision restricted to subgroup $g$ and "ego only" is the vehicle using its own sensors. We predict two things. First, $D_g(b)$ is smaller for assistive-device and child-scale pedestrians than for standard adults. Second, this gap widens as $b$ shrinks, because compression keeps large, dense structures and discards sparse thin ones.
The constructive claim is about where the extra bandwidth should go. Spending it on cells chosen by the receiver's own uncertainty about the person class recovers more of the subgroup gap than either uniform bandwidth or sender-chosen regions. Sender-chosen regions depend on the sender's detector already firing on an unusual shape, which is exactly what fails for these road users.
Grounding in Prior Work
Several papers motivate the gap. ProGait builds a video benchmark of transfemoral prosthesis users and shows that off-the-shelf segmentation and pose models fail on prosthetic limbs until they are fine-tuned. Fine-tuning recovers a pose AP of 0.942. Its parallel-bar condition also shows that physical structures hide the lower limbs. The concurrent-validity study on older adults tests pose trackers on 11 people with a mean age of 85. Temporal gait measures are nearly perfect (R = 0.99 to 1.00), but spatial measures such as step width collapse to R = 0.30 to 0.60 when one leg hides the other. HAR-PMD shows from wearable sensors that walking, crutches, walkers, and manual and electric wheelchairs produce distinct motion signatures, at 98.79% user-independent accuracy. It has no external cameras or LiDAR, so it says nothing about how these users look to a vehicle. Together these papers tell us that appearance and geometry really differ across groups. None of them involves occlusion by traffic or shared perception.
A second group of papers supplies the methods we reuse and constrains them. V2VNet shares compressed intermediate feature maps between vehicles through a graph network. It reports more than 60% higher AP on objects with almost no LiDAR returns, at 417× compression. OPV2V introduces attention-based intermediate fusion in a simulated CARLA benchmark and reports only about a 3% AP drop at 4096× compression. That robustness was measured on 232,913 vehicle boxes. Our hypothesis is that it does not transfer to thin human structures, and OPV2V is the result we must directly contradict or confirm. CoBEVT fuses multi-agent camera features in a top-down bird's-eye-view (BEV) grid with sparse attention. TUMTraf V2X fuses real roadside and vehicle features by max-pooling in BEV (CoopDet3D) and labels pedestrians as one class. DAIR-V2X quantifies the trade-offs between fusion styles on real intersections: late fusion reaches 41.90 AP, early fusion 50.03 AP, and the vehicle alone 31.33 AP. Its bandwidth figures are 0.3 KB for late fusion versus 1.3 MB for early fusion. EMIFF corrects camera misalignment caused by timing offsets. Every one of these papers reports only aggregate or vehicle-centric metrics.
The closest method is PACP. It allocates bandwidth across sending agents by solving a submodular selection over BEV overlap, gaining up to 13.60% AP over fairness-based allocation. PACP decides which agents to listen to, based on geometry. We decide which grid cells to refine and at what fidelity, based on the receiver's uncertainty about people. We also evaluate by subgroup rather than by total AP.
Research Program
The first component is the benchmark. No cooperative dataset in the file labels assistive-device pedestrians, so we build one in simulation. We extend OPV2V-style CARLA/OpenCDA scenes with pedestrian assets for wheelchair users, walker users, crutch users, and child-scale walkers, with realistic speeds and placements. We render the same LiDAR and camera suites, and we deliberately place pedestrians in occluded positions behind buses and corners. Simulation is a limitation: CARLA's stock rigs are the very homogeneity the file criticizes, and our inserted assets will be imperfect. We therefore also hand-annotate subgroup attributes on the pedestrian boxes of TUMTraf V2X and DAIR-V2X wherever such pedestrians appear. We expect those counts to be small and will report them honestly as a real-world sanity check rather than a headline number. All methods are trained on the same asset-augmented training split. This means that "more diverse data" is controlled for and the comparison isolates the communication scheme.
The second component is the measurement study. We sweep compression ratios from uncompressed through 64×, 417×, and 4096× (the ratios used by CoBEVT, V2VNet, and OPV2V) on attentive intermediate fusion, CoBEVT, and CoopDet3D-style max fusion. We report $D_g(b)$ for each subgroup and for each occlusion level, where occlusion level is measured as the fraction of the object's points visible from the ego vehicle.
The third component is the remedy, receiver-requested refinement. Every connected agent always sends a heavily compressed BEV map. The receiver fuses it and computes an uncertainty for each cell on its person-class output, using the entropy of the fused heatmap. It then requests lightly compressed feature patches for the most uncertain cells until a fixed byte budget is used up. The cost of the request round-trip is counted against both the latency budget and the byte budget. The key design choice is that the receiver chooses the cells, not the sender. The receiver can see a weak, ambiguous blob near the curb that no single sender would have scored highly.
Baselines
The naive reference point is ego-only detection with no sharing, trained on the same augmented data. It defines the zero of $D_g$. The communication-cheap reference is late fusion, in which each agent sends its own boxes. For it we report subgroup recall at the same bandwidth. The file's overall comparison is DAIR-V2X LiDAR late fusion at 41.90 AP versus 31.33 AP ego-only. We expect late fusion to lose the most on assistive-device users, because their boxes are dropped at the sender's confidence threshold.
The strongest existing intermediate-fusion methods are attentive fusion from OPV2V and CoBEVT on simulated data, and CoopDet3D on TUMTraf. The overall numbers in the file are 0.734 [email protected] for attentive fusion in default CARLA towns (OPV2V) and 90.76 mAP cooperative versus 76.40 vehicle-only for CoopDet3D (TUMTraf V2X). We found no reported result on Lacuna for any cooperative method on wheelchair, walker, crutch, or child pedestrian subgroups, so the subgroup numbers for these baselines will be produced here.
PACP is the strongest bandwidth-allocation competitor. We run it with the same backbone and the same total byte budget per frame.
Two ablations isolate the key ingredient. In the first, the extra bytes are spent uniformly across the map. In the second, the extra bytes go to regions the sender's detector already marks as pedestrian. Every method gets the same training data, backbone, number of tuning runs, and total bytes per frame.
The proposal counts as beating these baselines only if it does two things at equal bytes. It must cut the gap between the worst subgroup's dividend and the adult-walker dividend substantially, by at least half at 417× or higher compression. It must also keep overall AP within about one point of the best baseline. Improving average AP alone does not count.
Why not just…?
Why not just reweight the training loss toward rare pedestrian classes? Training the compressor and detector with higher weight on assistive-device classes is the cheapest fix, and it might close much of the gap. It is a real threat. We include it as a baseline applied to every fusion method, and we run our refinement on top of it. If reweighting alone brings $D_g(b)$ into line across subgroups at high compression, we report that as the finding. In that case the recommendation becomes a training change rather than a protocol change. Our reason to doubt it is that a fixed byte budget at 4096× may simply not have room for thin structures, however the loss is weighted.
Why not just send boxes (late fusion)? Late fusion is very cheap at about 0.3 KB per frame. But DAIR-V2X shows it trails early fusion by more than 8 AP even on aggregate classes. It also cannot recover an object the sender's detector suppressed, and ProGait suggests suppression is likely for unfamiliar shapes. Our subgroup recall comparison measures this directly.
Why not let the sender pick which regions to send in high fidelity? Sender-side selection is the second ablation above. It fails exactly when the sender does not recognize a walker or a wheelchair as a person. If it matches receiver-requested refinement, our mechanism adds nothing and we will say so.
Why not have vulnerable pedestrians broadcast their own presence from phones? HAR-PMD shows phones can classify mobility mode accurately. But phones give no 3D extent or reliable global position, and coverage among children or people with dementia cannot be assumed. Beacons can complement perception but cannot replace it.
Decisive Checks
Three outcomes decide the proposal. The first is whether the measured dividend actually differs across subgroups. If it is equal at every compression level, the premise fails and the OPV2V robustness result generalizes to people. The second is whether the subgroup gap grows with compression. The third is whether receiver-requested refinement beats both reweighting and sender-chosen refinement on the worst-subgroup dividend at matched bytes and latency. That comparison must hold in simulation and stay directionally consistent on the small set of annotated real-world TUMTraf and DAIR-V2X cases.
References
- ProGait: A Multi-Purpose Video Dataset and Benchmark for Transfemoral Prosthesis Users
- Concurrent validity of human pose tracking in video for measuring gait parameters in older adults: a preliminary analysis with multiple trackers, viewing angles, and walking directions
- Human Activity Recognition Dataset for Pedestrians with Mobility Disabilities
- V2VNet: Vehicle-to-Vehicle Communication for Joint Perception and Prediction
- OPV2V: An Open Benchmark Dataset and Fusion Pipeline for Perception with Vehicle-to-Vehicle Communication
- CoBEVT: Cooperative Bird’s Eye View Semantic Segmentation with Sparse Transformers
- TUMTraf V2X Cooperative Perception Dataset
- DAIR-V2X: A Large-Scale Dataset for Vehicle-Infrastructure Cooperative 3D Object Detection
- EMIFF: Enhanced Multi-scale Image Feature Fusion for Vehicle-Infrastructure Cooperative 3D Object Detection
- PACP: Priority-Aware Collaborative Perception for Connected and Autonomous Vehicles