Vision-Map Fusion Multi-Object Tracking at Complex Intersections Using HD Map Priors and Nonlinear Filtering

Abstract

Accurate multi-object tracking and metric localization support traffic monitoring and cooperative intelligent transportation at complex intersections. This study presents a fixed-camera vision-map fusion framework that addresses two practical difficulties: axis-aligned boxes poorly represent turning vehicles, and unconstrained image-plane tracking can produce physically implausible trajectories. A map-aided frontend first generates candidate detections using improved You Only Look Once version 8 nano (YOLOv8n) horizontal bounding box (HBB) branch and an improved YOLOv8 oriented bounding box (OBB) branch. A high-definition (HD) map selector then retains the candidate geometry consistent with the straight-driving or turning region and converts it into a unified detection record. The selected reference point is projected to the ground plane through an offline-estimated homography, whereas the appearance feature bypasses the homography and is passed directly to the association stage. The tracking backend uses a 12-dimensional joint image/metric state, symmetric central-difference evaluations of the process and measurement functions, appearance-motion association, and a feasible-road projection derived from HD-map lane polygons. On the evaluated public sequences, the complete configuration achieved a multiple object tracking accuracy (MOTA) of 74.5%, an identification F1 score (IDF1) of 82.6%, 614 identity switches, and a throughput of 26.8 frames per second (FPS) on an RTX 4090 workstation. In a descriptive Vehicle-in-the-Loop case study involving one instrumented vehicle at one intersection, the overall localization mean absolute error (MAE) was 0.180 m, compared with 0.208 m for the baseline end-to-end configuration. These results indicate the feasibility of combining branch-specific vehicle geometry with map-constrained tracking; controlled same-detector comparisons, repeated multi-vehicle trials, and embedded-device latency and power profiling remain necessary for broader claims.

1. Introduction

Intersections are critical nodes in urban traffic networks, where monitoring accuracy affects signal control, safety analysis, and cooperative intelligent transportation [1,2]. Their visual complexity arises from intersecting traffic streams, heterogeneous road users, frequent occlusion, illumination changes, and abrupt changes in vehicle heading. A roadside tracking system must therefore preserve target identity while also producing locations that remain meaningful in the road coordinate system.

Recent vision-based road-transport studies have adopted one-stage You Only Look Once (YOLO) detectors because they offer a useful accuracy-throughput compromise [3,4]. However, horizontal bounding boxes (HBBs) include increasing background area when a vehicle turns, whereas applying oriented bounding boxes (OBBs) over an entire intersection can increase computation. This motivates a spatially selective design in which the map determines where HBB or OBB geometry is used, rather than treating one representation as optimal for every part of the scene.

Tracking-by-detection methods such as Simple Online and Realtime Tracking (SORT), DeepSORT, and StrongSORT combine motion prediction with data association [5,6,7]. Recent work in Automation has also emphasized continuity under missed detections and posture-aware multi-target tracking [8,9]. At intersections, constant-velocity image-plane prediction is challenged by curved motion, acceleration, scale changes, and long occlusion. Moreover, a visually smooth track can still cross a curb or opposing lane because image-plane estimators do not encode the drivable topology.

To address these linked issues, this paper develops a fixed-camera vision-map fusion framework. The frontend generates HBB and OBB candidate detections and then uses the high-definition (HD) map regions, together with the detection reference point or predicted track point, to select the appropriate candidate representation. The backend converts the selected candidate into a unified observation, projects only its reference point into the metric plane, and passes the appearance feature directly to the association stage. It then performs central-difference state estimation and appearance–motion association and projects the posterior position to the nearest drivable-map location only when it lies outside the feasible road set. Unlike camera–radar fusion [10], the operational inputs remain a monocular roadside video stream and a pre-registered static map; the real-time kinematic (RTK) receiver is used only as an evaluation reference.

The contributions are summarized as follows:

• A unified vision-map tracking formulation is introduced in which image-plane and metric-plane variables are maintained in one state, and HD-map lane polygons provide an explicit feasibility correction for off-road predictions.
• A map-aided dual-box frontend combines an HBB branch enhanced with a cross-stage partial bottleneck with two convolutions and squeeze-and-excitation (C2fSE), dense atrous spatial pyramid pooling (DenseASPP), and Wise-IoU version 3 (WIoU v3) training loss with an OBB branch enhanced with spatial pyramid pooling-fast (SPPF), large separable kernel attention (LSKA), and DySample. Both branches first generate candidate detections. An HD-map selector subsequently retains the appropriate HBB or OBB representation and constructs a common detection record without reinitializing the tracker identity.
• Public-dataset detection and tracking results are complemented by a single-site RTK-referenced case study. The study reports the observed metric error and explicitly distinguishes this descriptive validation from multi-site or embedded-deployment verification.

2. Related Work

2.1. Limitations of HBB-Based Detection Algorithms

One-stage detectors, particularly YOLO variants, are widely used for road-object perception because of their speed and compact model families [3,4]. End-to-end transformer detectors provide another real-time design direction [11]. Recent studies in Sensors and Electronics have examined YOLO-based roadside vehicle detection, camera–radar fusion, and deployment on constrained edge devices [4,10]. These studies also show that workstation throughput cannot be transferred directly to embedded hardware without device-specific latency, memory, and power measurements [4]. For the straight-driving branch, the present work retains HBB geometry and targets occlusion and scale variation through channel recalibration, dense atrous context, and dynamic box-regression focusing.

OBB detection represents vehicle orientation more directly and is useful for turning or diagonally viewed vehicles. Region-of-interest (RoI) Transformer is an early representative approach for oriented detection [12], while Rope3D provides roadside orientation and 3D annotations for traffic perception [13]. DenseASPP aggregates context at several dilation rates [14], and DySample learns sampling offsets for lightweight upsampling [15]. Rather than treating either box representation as universally optimal, the proposed frontend generates HBB and OBB candidates and uses map-defined straight-driving and turning regions to select the appropriate candidate geometry before tracking.

2.2. Vision-Based Multi-Object Tracking and Localization

Tracking-by-detection associates frame-wise observations with predicted tracks. ByteTrack and ByteTrackV2 preserve low-confidence detections to improve association continuity [16,17], whereas DeepSORT and StrongSORT add appearance information and stronger association or motion components [5,6]. Recent approaches have also addressed missed detections or posture-specific tracking in other application domains [8,9]. The present study focuses on fixed-camera vehicle motion at intersections and evaluates a central-difference state estimator derived from symmetric evaluations of nonlinear process and measurement functions [18].

Metric localization from a fixed monocular camera requires an explicit scene model. For an approximately planar road surface, a homography can map image reference points to a ground-plane coordinate system under the standard projective-geometry model [19], with robust estimation based on feature or control-point correspondences [20,21]. MapTR and MapTRv2 instead learn online vectorized HD-map representations [22,23], while BEVFusion and BEVFormer construct bird’s-eye-view (BEV) representations from multiple sensors or multiple camera views [24,25]. The present method uses a pre-registered static map as prior information rather than a second live sensor. Camera-radar or camera–light detection and ranging (LiDAR) systems provide complementary measurements [10] but require additional sensing, synchronization, and calibration. The fixed-camera configuration reduces online sensor requirements but remains sensitive to camera motion, non-planar road geometry, adverse weather, and calibration drift.

3. Proposed Methodology

3.1. Problem Formulation and System Overview

This section defines the proposed fixed-camera multi-object tracking and localization framework. The online source is a calibrated roadside image sequence. The static map contains lane polygons, centerlines, and map-defined regions of interest. The homography ${\mathit{H}}_{\mathrm{map}}$ between image and local ground coordinates is estimated offline and is not recomputed for every frame.

The method is organized into five layers. Layer 1 acquires roadside video, map geometry, and offline camera-map control points. Layer 2 performs map-aided dual-box perception. Layer 3 converts the selected detection reference point from image coordinates to a common six-dimensional joint metric–image measurement vector. Layer 4 predicts, associates, updates, and, when required, projects the track onto the feasible road set. Layer 5 outputs track identities, image and metric trajectories, and evaluation metrics.

Figure 1 follows these five layers explicitly. The HBB and OBB branches first generate candidate detections. The map-aided selector evaluates the reference point of an unmatched detection or the predicted metric point of an active track against the straight-driving region ${\Omega }_S$ and turning region ${\Omega }_T$, and then retains the appropriate candidate representation. The selected candidate is converted into a unified record containing the box geometry and type, class, score, reference point, appearance feature, and, when available, the OBB angle $\theta$ as detection metadata. Only the reference point is projected through the homography ${\mathit{H}}_{\mathrm{map}}$ to form the six-dimensional joint metric–image measurement vector ${\mathit{z}}_k={\left(X_k^{\mathrm{obs}},Y_k^{\mathrm{obs}},u_k^{\mathrm{obs}},v_k^{\mathrm{obs}},b_{w,k}^{\mathrm{obs}},b_{h,k}^{\mathrm{obs}}\right)}^T$. The appearance feature does not belong to ${\mathit{z}}_k$ and therefore bypasses the homography and directly enters the appearance–motion association stage. The previous posterior state enters central difference Kalman filter (CDKF) prediction, while the predicted state and covariance, current six-dimensional joint metric–image measurement vector, and appearance feature jointly support association and measurement update. The resulting posterior position is corrected only when it leaves the HD-map feasible set.

Abbreviations: HD, high-definition; HBB, horizontal bounding box; OBB, oriented bounding box; C2f, cross-stage partial bottleneck with two convolutions; C2fSE, C2f with squeeze-and-excitation; DenseASPP, dense atrous spatial pyramid pooling; WIoU v3, Wise-IoU version 3; SPPF, spatial pyramid pooling-fast; LSKA, large separable kernel attention; CDKF, central difference Kalman filter; ID, identity; 2D, two-dimensional; MOTA, multiple object tracking accuracy; IDF1, identification F1 score; IDS, identity switches; FPS, frames per second; RTK, real-time kinematic.

3.2. Spatial-Decoupled Dual-Box Perception Frontend

The intersection ROI is partitioned into a straight-driving set ${\Omega }_S$ and a turning set ${\Omega }_T$ using HD-map lane polygons. Both detector branches first produce candidate detections. The selector then evaluates the predicted metric reference point of each active track and the reference point of each unmatched detection. Candidates associated with ${\Omega }_S$ retain HBB geometry, whereas candidates associated with ${\Omega }_T$ retain OBB geometry. A branch label is metadata attached to an observation rather than a tracker identity, and both representations share the same downstream observation schema and tracker pool.

For detections assigned to ${\Omega }_S$, the YOLOv8n HBB branch is used. Its three modifications have distinct roles.

• C2fSE channel recalibration: A squeeze-and-excitation (SE) [26] is embedded in the C2f path. The C2f output follows two paths: one produces channel weights and the other supplies the unweighted transformed feature to element-wise multiplication. The reweighted feature is then added to the unchanged shortcut feature. Figure 2 shows both inputs to the multiplication and residual-addition nodes explicitly.
• DenseASPP context aggregation: Densely connected atrous convolutions with dilation rates 1, 3, and 5 aggregate multiple receptive fields without an additional reduction in feature-map resolution [14]. This module targets the scale variation between small distant vehicles and large near-field vehicles.
• WIoU v3 regression loss: The non-monotonic focusing mechanism adjusts the contribution of anchors according to their outlier degree [27]. Its purpose is to reduce the dominance of very low-quality samples during box regression; it does not guarantee that every intermediate ablation will improve every metric.

For detections assigned to ${\Omega }_T$, the YOLOv8-OBB branch returns an oriented box parameterized by center coordinates, width, height, and heading angle. At the ${\Omega }_S$/${\Omega }_T$ boundary, the branch label may change between frames, but the metric reference point, appearance descriptor, and predicted track state are passed to the same association stage. The track identifier and covariance are not reset. If both ROIs overlap, candidate detections are transformed to the common space and the candidate with the lower gated association cost is retained, which reduces duplicate updates and identity fragmentation.

• SPPF-LSKA attention: LSKA [28] is embedded after spatial-pyramid pooling to capture long-range spatial dependencies using separable large-kernel operations. This is intended to retain elongated vehicle structure during turns without applying a dense large 2D kernel.
• DySample upsampling: The LSKA feature is supplied to a lightweight offset predictor to estimate the sampling offset $\Delta G$. The regular grid G does not enter the offset predictor; instead, $G$ and $\Delta G$ are combined only at the addition node to form $G^{\prime }=G+\Delta G$. The LSKA feature is then sampled at $G^{\prime }$, and the resulting feature is decoded into OBB confidence, category, center, size, and angle [15].

Figure 2 links the perception frontend to the tracking backend at the data-flow level. Panel A shows that the C2f output U supplies both the SE-weight generation path and the feature-reweighting path. In the DySample block, the LSKA feature predicts $\Delta G$, the regular grid $G$ bypasses the offset predictor, and $G$ and $\Delta G$ are added to form $G^{\prime }$ before feature sampling. Panel B separates process prediction from measurement processing. The previous posterior generates the predicted state and covariance, which are supplied both to the measurement-function evaluations and directly to appearance–motion association for motion gating. The current six-dimensional joint metric–image measurement vector and appearance feature enter the association and measurement-update stage through separate paths; the appearance feature is not included in ${\mathit{z}}_k$ and is not transformed by the homography. The implemented map operation remains a direct nearest-feasible projection rather than an iterative quadratic-programming loop.

3.3. Map-Constrained CDKF Tracking and Localization Backend

The backend extends a DeepSORT-style tracking-by-detection pipeline with a joint image/metric state and a map-feasibility operation. The estimator is referred to as a Central Difference Kalman Filter (CDKF) because local process and measurement sensitivities are obtained from symmetric function evaluations rather than analytic Jacobians [18].

Let the observed image reference point and its corresponding local ground-plane point be related through the standard planar projective model [19], as expressed in Equation (1). The image-to-ground homography ${\mathit{H}}_{\mathrm{map}}$ is estimated offline from paired camera-map control points using scale-invariant feature transform (SIFT) descriptors [21] and random sample consensus (RANSAC) [20]. Equation (2) gives the dehomogenized metric coordinates. The HBB bottom center and the OBB ground-contact reference point are both converted by this mapping. Because the planar model is a 2D-to-2D projective mapping, the output is a metric ground-plane observation rather than a full 3D reconstruction.

$${\lambda }_k\left[\begin{array}{c}{X}_k^{\mathrm{obs}}\\ Y_k^{\mathrm{obs}}\\ 1\end{array}\right]={\mathit{H}}_{\mathrm{map}}\left[\begin{array}{c}{u}_k^{\mathrm{obs}}\\ v_k^{\mathrm{obs}}\\ 1\end{array}\right],\hspace{1em}\hspace{1em}{\mathit{H}}_{\mathrm{map}}={\left[h_{ij}^{\mathrm{map}}\right]}_{3\times 3}$$

(1)

$$\begin{array}{l}{X}_k^{\mathrm{obs}}={\displaystyle \frac{h_{11}^{\mathrm{map}}{u}_k^{\mathrm{obs}}+h_{12}^{\mathrm{map}}{v}_k^{\mathrm{obs}}+h_{13}^{\mathrm{map}}}{h_{31}^{\mathrm{map}}{u}_k^{\mathrm{obs}}+h_{32}^{\mathrm{map}}{v}_k^{\mathrm{obs}}+h_{33}^{\mathrm{map}}}},\\ Y_k^{\mathrm{obs}}={\displaystyle \frac{h_{21}^{\mathrm{map}}{u}_k^{\mathrm{obs}}+h_{22}^{\mathrm{map}}{v}_k^{\mathrm{obs}}+h_{23}^{\mathrm{map}}}{h_{31}^{\mathrm{map}}{u}_k^{\mathrm{obs}}+h_{32}^{\mathrm{map}}{v}_k^{\mathrm{obs}}+h_{33}^{\mathrm{map}}}}.\end{array}$$

(2)

In Equations (1) and (2), $k$ denotes the time-step index; $(u_k^{\mathrm{obs}},v_k^{\mathrm{obs}})$ denotes the observed image-plane reference point in pixels; and $(X_k^{\mathrm{obs}},Y_k^{\mathrm{obs}})$ denotes the corresponding observed ground-plane position in meters. The non-zero scalar ${\lambda }_k$ is the homogeneous scale factor. The matrix ${\mathit{H}}_{\mathrm{map}}\in {ℝ}^{3\times 3}$ is the image-to-ground homography, and $h_{ij}^{\mathrm{map}},i,j\in \{1,2,3\}$, denotes its element in the i-th row and j-th column. The superscript $\mathrm{obs}$ identifies a measured quantity, and the superscript $T$ denotes matrix transpose.

At time step $k$, the 12-dimensional state of each active track contains metric position, metric velocity and acceleration, image-plane position and velocity, and bounding-box dimensions. The six-dimensional measurement contains the observed metric position, image-plane reference point, and bounding-box dimensions, as defined in Equation (3).

$$\begin{array}{l}{\mathit{x}}_k={\left(X_k,Y_k,V_{X,k},V_{Y,k},a_{X,k},a_{Y,k},u_k,v_k,V_{u,k},V_{v,k},b_{w,k},b_{h,k}\right)}^T,\\ {\mathit{z}}_k={\left(X_k^{\mathrm{obs}},Y_k^{\mathrm{obs}},u_k^{\mathrm{obs}},v_k^{\mathrm{obs}},b_{w,k}^{\mathrm{obs}},b_{h,k}^{\mathrm{obs}}\right)}^T.\end{array}$$

(3)

In Equation (3), ${\mathit{x}}_k\in {ℝ}^{12}$ is the state vector of an active track at time step $k$, and ${\mathit{z}}_k\in {ℝ}^6$ is the corresponding measurement vector. The variables $X_k$ and $Y_k$ are the metric ground-plane coordinates in meters; $V_{X,k}$ and $V_{Y,k}$ are the metric velocities in meters per second; and $a_{X,k}$ and $a_{Y,k}$ are the metric accelerations in meters per second squared. The variables $u_k$ and $v_k$ are the image-plane coordinates in pixels, whereas $V_{u,k}$ and $V_{v,k}$ are their image-plane velocities in pixels per second. The variables $b_{w,k}$ and $b_{h,k}$ denote the bounding-box width and height in pixels. The prefix $b$ distinguishes the bounding-box dimensions from the process and measurement functions.

Equation (4) defines the stochastic process and measurement models. The function f propagates the metric components with a constant-acceleration model and the image components with a constant-velocity model. The process noise has covariance Q. The function h selects or nonlinearly maps the predicted state into the six measured quantities, and the measurement noise has covariance R.

$$\begin{array}{l}{\mathit{x}}_k=\mathit{f}\left({\mathit{x}}_{k-1},\Delta t\right)+{\mathit{q}}_{k-1},{\mathit{q}}_{k-1}~\mathcal{N}\left(0,\mathit{Q}\right),\\ {\mathit{z}}_k=\mathit{h}\left({\mathit{x}}_k\right)+{\mathit{r}}_k,{\mathit{r}}_k~\mathcal{N}\left(0,\mathit{R}\right).\end{array}$$

(4)

In Equation (4), $\mathit{f}:{ℝ}^{12}\times {ℝ}_{+}\to {ℝ}^{12}$ is the state-transition function, $\mathit{h}:{ℝ}^{12}\to {ℝ}^6$ is the measurement function, and $\Delta t$ is the elapsed time between two consecutive frames in seconds. The vectors ${\mathit{q}}_{k-1}$ and ${\mathit{r}}_k$ are the zero-mean process and measurement noises, respectively. Their covariance matrices are $\mathit{Q}\in {ℝ}^{12\times 12}$ and $\mathit{R}\in {ℝ}^{6\times 6}$. The notation $\mathcal{N}(0,\mathit{Q})$ denotes a Gaussian distribution with zero mean and covariance $\mathit{Q}$, with an analogous definition for $\mathcal{N}\left(0,\mathit{R}\right)$.

$$\begin{array}{l}{\left({\mathit{J}}_f\right)}_{:,j}\approx {\displaystyle \frac{\mathit{f}\left({\widehat{\mathit{x}}}_{k-1|k-1}+{\delta }_j{\mathit{e}}_j,\Delta t\right)-\mathit{f}\left({\widehat{\mathit{x}}}_{k-1|k-1}-{\delta }_j{\mathit{e}}_j,\Delta t\right)}{2{\delta }_j}},\\ {\left({\mathit{J}}_h\right)}_{:,j}\approx {\displaystyle \frac{\mathit{h}\left({\widehat{\mathit{x}}}_{k|k-1}+{\delta }_j{\mathit{e}}_j\right)-\mathit{h}\left({\widehat{\mathit{x}}}_{k|k-1}-{\delta }_j{\mathit{e}}_j\right)}{2{\delta }_j}},\hspace{1em}\hspace{1em}j=1,\dots ,12\end{array}$$

(5)

.

In Equation (5), ${\mathit{J}}_f\in {ℝ}^{12\times 12}$ and ${\mathit{J}}_h\in {ℝ}^{6\times 12}$ denote the process and measurement sensitivity matrices, respectively, and ${(\cdot )}_{:,j}$ denotes the j-th matrix column. The vector ${\mathit{e}}_j\in {ℝ}^{12}$ is the j-th unit basis vector, and ${\delta }_j>0$ is the central-difference perturbation applied to the j-th state component. The estimate ${\widehat{\mathit{x}}}_{k-1|k-1}$ is the posterior state at the previous time step, whereas ${\widehat{\mathit{x}}}_{k|k-1}$ is the current prior state. More generally, the notation $k/l$ denotes an estimate at time $k$ conditioned on observations available through time $l$. The matrices ${\mathit{P}}_{k-1|k-1},{\mathit{P}}_{k|k-1}$, and ${\mathit{P}}_{k|k}$ denote the previous posterior, current prior, and current posterior error covariances, respectively. The current observation ${\mathit{z}}_k$ is not used as an input to state prediction; it enters only the association and measurement-update stage.

Let ${\widehat{\mathit{p}}}_{k|k}={[{\widehat{X}}_{k|k},{\widehat{Y}}_{k|k}]}^T$ denote the unconstrained posterior metric position extracted from ${\widehat{\mathit{x}}}_{k|k}$, and let $\mathcal{M}$ denote the union of legally drivable lane polygons. The map-feasibility operation is defined in Equation (6).

$${\widehat{\mathit{p}}}_{k|k}^c=\left\{\begin{array}{ll}{\widehat{\mathit{p}}}_{k|k},\hfill & {\widehat{\mathit{p}}}_{k|k}\in \mathcal{M},\hfill \\ {\Pi }_{\mathcal{M}}\left({\widehat{\mathit{p}}}_{k|k}\right)\equiv \underset{\mathit{p}\in \mathcal{M}}{\mathit{arg}\mathit{min}}{‖\mathit{p}-{\widehat{\mathit{p}}}_{k|k}‖}_2,\hfill & {\widehat{\mathit{p}}}_{k|k}\notin \mathcal{M}.\hfill \end{array}\right.$$

(6)

In Equation (6), the superscript $c$ denotes a map-corrected quantity, ${\Pi }_{\mathcal{M}}(\cdot )$ is the Euclidean projection onto the feasible set $\mathcal{M}$, and $\Vert \cdot {\Vert }_2$ denotes the Euclidean norm. When the posterior position lies outside $\mathcal{M}$, only its metric position components are replaced by the nearest feasible point; track identity, appearance history, and covariance history are retained. This deterministic geometric projection is triggered only for an off-road posterior.

4. Experiments, Results, and Discussion

4.1. Experimental Setup and Implementation Details

Experiments used the official partitions of four public datasets. The University at Albany Detection and Tracking (UA-DETRAC) benchmark [29] contains 100 traffic videos and more than 140,000 annotated frames at 960 × 540 resolution; VisDrone2019 [30] contains 6471 training, 548 validation, and 1580 test images; the Berkeley DeepDrive 100K (BDD100K) dataset [31] contains 100,000 road images; and Rope3D [13] contains approximately 50,000 roadside images with more than 1.5 million 3D annotations, converted to the OBB representation used here. The HBB and OBB configurations contain 3.19 million and 3.23 million parameters and require 8.5 and 9.1 giga floating-point operations (GFLOPs) per image, respectively. All compared detection variants were trained from scratch under the same augmentation policy, with an initial learning rate of 0.02 and momentum of 0.9. Random flipping and mosaic augmentation were used. The reported 26.8 frames per second (FPS) end-to-end result was measured on an NVIDIA RTX 4090 desktop workstation; it is not an embedded-device benchmark. The physical deployment used a Jetson AGX Orin 32 GB node for acquisition and online processing. Standardized per-stage latency, peak memory, and power measurements are not reported; therefore, embedded-device deployment efficiency is not quantified in this study and remains a subject for future evaluation [4,32].

4.2. Ablation Studies on Dual-Box Perception

Ablations start from the corresponding YOLOv8 baseline and add the modules in the order shown in

Table 1. Ablation experiments of the HBB branch in the straight-driving region (evaluated on the UA-DETRAC dataset).

Model	C2fSE	DENSEASPP	WIoU v3	Precision (%)	Recall (%)	mAP@0.5 (%)	mAP0.5–0.95 (%)
Baseline				96.7	96.3	98.9	88.6
+C2fSE	✓			97.5	96.7	98.7	88.8
+DenseASPP	✓	✓		96.1	96.9	98.8	88.3
Proposed	✓	✓	✓	96.9	98.0	99.1	89.1

and

Table 2. Sequential ablation of the OBB branch on Rope3D; intermediate rows need not improve every metric.

Model	LSKA	DySample	Precision (%)	Recall (%)	mAP@0.5 (%)	mAP0.5–0.95 (%)
Baseline			82.9	66.8	74.9	41.7
+LSKA	✓		84.0	66.1	74.7	40.7
Proposed	✓	✓	85.6	66.8	75.6	42.7

. These are sequential configurations rather than independent single-factor tests, so an intermediate row can improve one metric while reducing another.

Table 1 shows a non-monotonic HBB trend. Adding C2fSE increases precision from 96.7% to 97.5% and recall from 96.3% to 96.7%, but mean average precision at an intersection-over-union threshold of 0.5 (mAP@0.5) changes from 98.9% to 98.7%. After DenseASPP, recall rises to 96.9% while precision and mean average precision averaged over intersection-over-union thresholds from 0.5 to 0.95 (mAP@0.5–0.95) decrease to 96.1% and 88.3%. These small reversals are consistent with a precision–recall and localization trade-off: channel reweighting and larger receptive fields alter which difficult or partially occluded boxes receive high confidence. WIoU v3 then changes the regression weighting, and the complete configuration reaches 99.1% mAP@0.5 and 89.1% mAP@0.5–0.95. The table therefore supports the final combination, not a claim that each addition monotonically improves every metric.

Table 2 exhibits the same trade-off in the OBB branch. LSKA raises precision from 82.9% to 84.0%, while recall decreases from 66.8% to 66.1% and mAP@0.5–0.95 changes from 41.7% to 40.7%. A plausible mechanism is that stronger long-range attention suppresses ambiguous positives and produces a more conservative score distribution. After DySample is added, precision, recall, mAP@0.5, and mAP@0.5–0.95 reach 85.6%, 66.8%, 75.6%, and 42.7%, respectively. Because the ablation is sequential, these results should be interpreted as evidence for the complete branch configuration rather than isolated causal effects.

Figure 3 provides representative qualitative comparisons for turning, dense-traffic, and nighttime scenes. In the selected examples, the revised branch produces boxes that better follow inclined vehicles and retain more detections in crowded regions. These examples illustrate typical behavior but do not replace the quantitative tables; additional comparisons are provided in Appendix B.

4.3. Baseline Comparisons on Multi-Object Tracking

Table 3. End-to-end tracking comparison. Detector and association configurations were not fully harmonized; values should not be interpreted as a filter-only ablation.

Tracker	Tracking Backend	MOTA(↑)	MOTP(↑)	IDF1(↑)	IDS(↓)	FPS
SORT	Linear KF	61.4%	72.1%	71.1%	1421	39.6
DeepSORT	Linear KF	70.8%	78.3%	78.4%	784	25.2
StrongSORT	Linear KF	70.4%	78.2%	81.2%	674	13.4
Proposed	CDKF + Map Prior	74.5%	80.2%	82.6%	614	26.8

compares complete tracking configurations on the evaluated intersection sequences. SORT, DeepSORT, StrongSORT, and the proposed system were not rerun with an identical frozen detection file and fully harmonized association settings. The table is therefore an end-to-end system comparison, not a controlled isolation of the CDKF and map prior.

Within this system-level comparison, the proposed configuration obtains a multiple object tracking accuracy (MOTA) of 74.5%, a multiple object tracking precision (MOTP) of 80.2%, an identification F1 score (IDF1) of 82.6%, 614 identity switches, and 26.8 FPS. StrongSORT has fewer identity switches than DeepSORT but lower throughput, while SORT is faster but less accurate. These results show a favorable accuracy–continuity–throughput balance for the full proposed pipeline. They do not establish that the difference is caused solely by the filter or map constraint; a future study should use identical detections and association thresholds and ablate CDKF and map projection separately.

4.4. Single-Intersection Validation Against an RTK Reference

The physical evaluation was conducted at one signalized urban intersection in Hangzhou, China, using one instrumented vehicle. It was designed as a Vehicle-in-the-Loop case study to check whether image-to-map trajectories were consistent with an external RTK reference. Because it covers one site and one vehicle run, the reported error is a descriptive result rather than a population estimate.

Figure 4 summarizes the instrumentation. The roadside unit used a Dahua DH-CP902-YGSL-L traffic camera with a 1.1-inch global-shutter complementary metal–oxide–semiconductor (GS-CMOS) sensor, 4096 × 2336 maximum resolution, a 25 FPS default stream (up to 50 FPS), and 100 dB wide dynamic range. The field node was an NVIDIA Jetson AGX Orin with 32 GB memory under Ubuntu 20.04 with Robot Operating System 2 (ROS 2), distinct from the RTX 4090 workstation used for the 26.8 FPS benchmark. A roof-mounted antenna and Beiyun X1 terminal recorded the reference trajectory at 10 Hz. The camera and RTK streams were recorded through ROS 2 and reviewed on the synchronized Foxglove playback timeline. The original auxiliary temporal-pairing script and the raw per-sample camera/RTK timestamp and residual files were not retained. Consequently, the acquisition-time clock specification, timestamp-offset correction, temporal-pairing rule, maximum accepted time gap, and final matched-sample count cannot be verified retrospectively. To support methodological replication without overstating the provenance of the field-trial results, a reconstructed and tested reference implementation is provided in the public release v1.0.0 at https://github.com/975032368ma-ai/vision-map-tracking-reproducibility/releases/tag/v1.0.0 (accessed on 12 August 2026). Appendix A summarizes the available configuration records, the scope of the released implementation, and the remaining limitations.

Figure 5 compares the reconstructed and RTK-referenced trajectories for representative straight and turning portions of the recorded run. In the displayed straight segments, both configurations remain close to the reference. In the displayed turning segments, the baseline trajectory shows a larger lateral departure, whereas the map-constrained trajectory remains inside the drivable polygon. This is a qualitative observation from the illustrated run and should not be generalized to untested intersections.

Figure 6 plots the lateral-position trajectory produced by the proposed vision-based localization pipeline together with the Beiyun RTK reference over the displayed 0–80 m trajectory segment.

Table 4. Descriptive localization MAE for the single-vehicle, single-intersection case study; configurations differ in multiple components (m).

Method	Box Representation	Spatial Constraint	Overall MAE (m)
Baseline end-to-end configuration	HBB	Unconstrained 2D	0.208
Proposed end-to-end configuration	OBB	HD-map feasible-set projection	0.180

separately reports aggregate mean absolute errors (MAEs) of 0.208 m for the HBB/unconstrained baseline and 0.180 m for the OBB/map-constrained configuration. Because the table changes box representation, filter, and spatial constraint together, it does not isolate the contribution of one component. Raw time-aligned residuals, the effective sample count N, and repeated trials were not reported in this study; consequently, no confidence interval or significance claim is reported.

4.5. Discussion, Computational Scope, and Limitations

The proposed approach occupies a middle ground between image-only tracking and live multi-sensor fusion. Camera-radar intersection tracking can supply complementary range or velocity measurements [10], whereas the present system uses a static HD map to regularize a monocular camera trajectory. This reduces online sensing hardware but cannot recover observations that are absent because of heavy rain, fog, glare, or complete visual occlusion.

Relative to recent automation studies, Predictive Enhancement of Detection (PrED) targets intermittent detections using prediction and similarity cues [8], while posture-aware DeepSORT improves tracking in a different occlusion-intensive domain [9]. The present contribution is distinct in its map-selected HBB/OBB geometry and explicit metric-road feasibility operation. The evidence is complementary rather than directly rankable because datasets, detectors, and target domains differ.

The reported 26.8 FPS is an aggregate workstation result. Offline homography calibration is excluded from per-frame processing, and the map correction is a direct point-in-polygon test followed, when necessary, by nearest-boundary projection rather than an iterative quadratic programming (QP) procedure. Nevertheless, reliable module-level timing was not recorded. Edge deployment studies emphasize that throughput, memory, thermal behavior, and power must be measured on the target device [4,32]; accordingly, no claim of verified Jetson real-time performance is made here.

Uncertainty is represented by the estimator covariance $Q/R$ model, but the experiments do not calibrate predictive uncertainty. Deep Gaussian-process work illustrates how uncertainty estimation can become overconfident and motivates future calibration studies [33]. A controlled follow-up should archive frame-level residuals, report confidence intervals, harmonize detections across trackers, and measure each pipeline stage on the deployed edge node.

Appendix A summarizes the RTK update rate, ROS 2/Foxglove workflow, coordinate frame, rotation from the body frame to the vehicle frame (RBV) procedure, and the map-origin and ROI parameters used in the field trial. It also identifies the calibration and synchronization quantities that could not be quantified in this study.

5. Conclusions

This study presented a fixed-camera vision-map fusion framework for vehicle detection, tracking, and metric localization at complex intersections. The method combines map-selected HBB/OBB perception, homographic image-to-ground projection, a central-difference joint-state estimator, appearance-motion association, and a nearest-feasible-road correction.

The final HBB and OBB configurations improved the selected detection metrics relative to their corresponding baselines, although intermediate ablations were non-monotonic. In the reported end-to-end tracking comparison, the proposed system achieved MOTA of 74.5%, IDF1 of 82.6%, 614 identity switches, and 26.8 FPS on an RTX 4090 workstation. Because the baseline trackers did not share fully harmonized detections and settings, these differences cannot be attributed solely to CDKF or map constraints.

In the single-vehicle, single-intersection RTK case study, the observed localization MAE was 0.180 m for the proposed configuration and 0.208 m for the baseline configuration. This comparison supports feasibility in the recorded case but does not demonstrate universal sub-meter accuracy, statistical superiority, or robustness across sites.

The main limitations are the single-site physical evaluation, absence of repeated multi-vehicle trials and confidence intervals, lack of a fixed-detection filter ablation, and absence of per-stage latency, memory, and power measurements on the Jetson device. The fixed-camera planar homography also assumes a stable camera and locally planar road surface.

Future work will collect synchronized per-frame residuals over multiple intersections and vehicles, report uncertainty intervals, and profile detection, projection, association, state estimation, and map correction separately on embedded hardware. Radar or LiDAR can be introduced as a true complementary sensing modality for adverse weather and long occlusion [10], while predictive uncertainty models such as deep Gaussian processes may inform adaptive process and measurement covariance design [33].