NHT-1071: Urban Non-Lane Traffic Dataset

• NHT-1071 is a dataset featuring 1071 images with approximately 30,000 annotated objects, capturing mixed traffic flows and directional movements in complex urban intersections.
• It employs a detailed annotation protocol with directional labels, optimized for object detection using YOLO models on low-resource devices for real-time traffic signal control.
• The dataset supports research in heterogeneous traffic modeling, trajectory prediction, and intelligent signal control in cities where conventional lane-based models fail.

The Non-lane-based and Heterogeneous Traffic (NHT-1071) dataset is developed to address the complexities of urban traffic in developing regions, where lane discipline is weak or absent, and traffic composition is highly varied. It is specifically curated for applications such as trajectory prediction, traffic signal optimization, and safety analysis in mixed and unstructured environments. NHT-1071 consists of annotated images featuring both motorized and non-motorized vehicles, explicitly segmented by directional movement into and out of intersections. The dataset is tailored for dense traffic contexts typified by Dhaka City, representing typical challenges faced in Asian urban environments where traditional lane-based models fail to generalize.

1. Dataset Structure and Composition

NHT-1071 contains 1071 images with approximately 30,000 annotated objects (Azam et al., 18 Oct 2025). Each instance belongs to one of four classes:

• Motorized vehicles entering the intersection
• Non-motorized vehicles entering the intersection
• Motorized vehicles exiting the intersection
• Non-motorized vehicles exiting the intersection

Class labels include directional information, which allows for fine-grained analysis of intersection dynamics and flow patterns. The data is collected under real-world conditions at high-density urban intersections, capturing scenarios typical of Dhaka's complex traffic: mixed vehicle types (cars, vans, rickshaws, cycles), no strict lane adherence, and frequent interactions between vehicles and pedestrians.

Image Count	Object Annotations	Classes	Direction info
1071	~30,000	4	Yes

This focus on both vehicle class and movement direction distinguishes NHT-1071 from lane-driven datasets such as KITTI or NGSIM, enabling research adapted to the requirements of urban South Asian traffic.

2. Annotation Protocol and Object Detection Frameworks

Annotation in NHT-1071 is optimized for object detection tasks using models such as YOLOv8 to YOLOv10 and RT-DETR, with bounding boxes demarcating each object and direction (Azam et al., 18 Oct 2025). The dataset was explicitly used to train and evaluate the YOLO family of detectors on resource-constrained devices (Raspberry Pi 4B with NCNN deployment), ensuring practical applicability for real-time systems.

Performance metrics on NHT-1071 were assessed using mean Average Precision (mAP) at IoU thresholds of 0.5 (mAP@50) and 0.5–0.95 (mAP@50–95), as well as inference time and model size. These evaluation metrics provide benchmarks for model selection and optimization. For Dhaka’s environment, lightweight YOLO variants (YOLOv9 via NCNN) were preferred due to their speed-accuracy trade-off when deployed on low-resource hardware.

Detector Variant	Deployment Platform	mAP@50	mAP@50–95	Inference Speed	Model Size
YOLOv9-NCNN	Raspberry Pi 4B	High	Moderate	Fast	Small

A plausible implication is that NHT-1071’s annotation protocol is designed to facilitate immediate downstream integration with embedded systems for real-time urban traffic monitoring.

3. Relevance in Non-lane-based and Heterogeneous Traffic Modeling

NHT-1071 explicitly targets settings where traffic discipline is weak and heterogeneous flows dominate. Unlike conventional datasets tailored to homogeneous, lane-structured traffic, its utility is maximized for algorithms and frameworks that model or predict phenomena such as:

• Mixed trajectory forecasting (vehicles with different dynamics sharing unmarked space)
• Intelligent signal control without assuming fixed lane positions
• Behavior segmentation in settings with frequent lane violations and irregular maneuvers

For example, YOLO-based detection on NHT-1071 enables adaptive traffic signal control at intersections by tracking vehicle flow and type in real time—critical for congestion optimization under non-lane-based traffic (Azam et al., 18 Oct 2025). Furthermore, models leveraging the dataset can adapt to intersection ingress/egress patterns by class, thus informing policy and infrastructure design that takes modality and directionality into account.

4. Integration with Advanced Signal Optimization Systems

NHT-1071 was implemented in a real-time intelligent traffic signaling testbed at the Palashi intersection in Dhaka (Azam et al., 18 Oct 2025). Live RTSP videos were processed on low-power edge devices, with detection outputs fed to a multi-objective optimization algorithm (NSGA-II). The objectives were formalized as:

$$\begin{align*} &f_1 = \sum_{i=1}^L (C_m(i) + C_{nm}(i)) \ &f_2 = \sum_{i=1}^L R_t(i) \end{align*}$$

with $L$ as the number of links (entry/exit roads), $C_m(i)$ and $C_{nm}(i)$ as vehicle counts, and $R_t(i)$ as red signal duration for each link.

The NSGA-II algorithm generates Pareto-optimal timing plans by iteratively evolving candidate solutions for minimizing vehicle waiting time and maximizing throughput, supporting multi-modal, non-lane-based flow management. The on-site deployment demonstrated the feasibility of embedding NHT-1071-detected vehicle flows into optimization pipelines while maintaining real-time performance on constrained hardware (average latency $\approx$ 8 seconds per cycle).

5. Comparative Context: Conventional Datasets and Urban Asian Traffic

Compared to datasets such as KITTI, NGSIM, or METEOR (Chandra et al., 2021), NHT-1071’s unique contribution is its tailored representation for mixed urban Asian intersections. Conventional datasets typically assume lane discipline and fewer agent categories, limiting their generalizability to environments like Dhaka or Hyderabad, where road users span motorized, non-motorized, and pedestrian classes and lanes are not consistently respected. NHT-1071’s annotation scheme—categorizing both vehicle type and movement direction—enables detection and optimization systems to operate under the spatial and behavioral diversity that defines non-lane-based traffic.

Dataset	Geographic Focus	Lane Discipline	Agent Types	Application Domains
KITTI	Europe/North Am.	Strict	~5	Lane-based detection, tracking
METEOR	India	Weak	16	Unstructured traffic, rare events
NHT-1071	Bangladesh	Weak	4 (+dir)	Signal control, detection, urban ITS

This suggests adoption of NHT-1071 may improve intersection management and research outcomes in cities with comparable traffic systems.

6. Significance for Intelligent Traffic Signaling and Urban Mobility

Integration of NHT-1071 into intelligent traffic signaling systems demonstrates substantial potential for context-sensitive congestion management (Azam et al., 18 Oct 2025). By enabling real-time classification and directional detection of motorized and non-motorized flows, the dataset permits fine-grained control over traffic light cycles in environments where the classical notion of “lane flow” is inapplicable.

The pipeline supports low-cost, scalable deployment in resource-constrained settings, with direct relevance to policy-makers and urban planners seeking adaptive IT solutions for developing cities. On-road evaluations at the Palashi intersection showed reductions in vehicle waiting counts when compared against manual police signaling, highlighting the dataset’s applicability to practical congestion reduction strategies.

A plausible implication is that NHT-1071, through its focus on mixed, directional traffic under non-lane conditions, is a pivotal resource for the development and benchmarking of robust ITS systems in rapidly urbanizing regions.

---

In summary, NHT-1071 is a purpose-built dataset supporting detection, optimization, and control of heterogeneous, non-lane-based urban traffic. Its distinctive annotation approach and proven integration into real-time signaling systems in Dhaka establish its value for research and deployment in complex urban environments characterized by high density, mixed modalities, and minimal lane discipline.