DriveIndia: Nexus for Indian Mobility Research

• DriveIndia is a research framework that unifies diverse Indian mobility datasets, emphasizing local traffic heterogeneity and unconstrained road conditions.
• It encompasses complementary 2D and 3D perception benchmarks that address domain adaptation challenges and open-set segmentation for complex environments.
• The agenda extends to ride-hailing analytics and Logistics 4.0, linking autonomous navigation insights with modern supply-chain and service design strategies.

DriveIndia denotes a set of India-specific mobility, perception, and logistics constructs that appear across multiple research contexts rather than a single universally fixed artifact. In the cited literature, the term is used for a large-scale object-detection dataset for Indian traffic scenes (Kumar et al., 26 Jul 2025), as an editorial rebranding of the Indian Driving Dataset (IDD) for autonomous navigation in unconstrained environments (Varma et al., 2018), and as a blueprint for Industry 4.0 and Logistics 4.0 modernization of Indian logistics infrastructure (Shriyam et al., 2023). Related work further connects this label to ride-hailing adoption analytics (Bhaduri et al., 2022), 3D perception in unstructured road scenes through IDD-3D (Dokania et al., 2022), unsupervised domain adaptation for semantic segmentation on IDD (Kothandaraman et al., 2020), and pothole-oriented driver assistance (Dwivedi et al., 2020). Taken together, these works position DriveIndia as a research nexus for Indian road-scene understanding, mobility-service design, and logistics digitalization under conditions of heterogeneous traffic, weak lane discipline, mixed infrastructure quality, and pronounced domain shift from benchmarks centered on structured urban roads.

1. Terminological scope and research context

A central feature of the DriveIndia literature is its emphasis on Indian operating conditions as a first-class modeling constraint. The IDD work defines the problem as autonomous navigation in “unconstrained environments,” where common assumptions of structured driving datasets do not hold: well-delineated lanes, low category variation, limited background diversity, and strict traffic-rule adherence are all weakened or absent (Varma et al., 2018). The later object-detection dataset named “DriveIndia” extends this emphasis by targeting “the complexity and unpredictability of Indian traffic environments,” with varied weather, illumination changes, heterogeneous road infrastructure, and dense mixed traffic patterns (Kumar et al., 26 Jul 2025).

This shared framing recurs outside perception. The logistics position paper uses “DriveIndia” to denote a layered modernization strategy built around Smart Industrial Parks (SIPs), IoT-based cyber-physical systems (CPS), Digital Twins (DTs), and policy incentives for adoption at scale (Shriyam et al., 2023). The ride-hailing study does not define a formal DriveIndia platform, but its end-to-end account explicitly describes how DriveIndia can translate latent user segments and attribute rankings into segment-tailored engagement and product design (Bhaduri et al., 2022). This suggests that “DriveIndia” functions as a unifying editorial and applied label for Indian mobility intelligence, spanning datasets, analytic workflows, and deployment blueprints.

A recurring misconception is that Indian mobility can be handled by straightforward transfer from Euro-American benchmarks and operating assumptions. The cited works argue otherwise through empirical domain gaps, new class definitions, long-tail object distributions, and specialized modeling pipelines for road and logistics environments (Varma et al., 2018, Pourhassan et al., 2023).

2. IDD and the formalization of unconstrained road-scene understanding

The foundational benchmark in this lineage is IDD, introduced as a dataset for exploring problems of autonomous navigation in unconstrained environments (Varma et al., 2018). It consists of 10,004 finely annotated images collected from 182 drive sequences on Indian roads. The data were captured with a calibrated stereo-camera rig mounted on a passenger car, using a forward-facing pair at 1678×968 px with no barrel distortion, and sampled more densely around intersections, market areas, and heavy junctions. Entire drive sequences were assigned wholesale to train, validation, or test sets to avoid scene leakage.

IDD defines 34 semantic classes, with 30 appearing in the level 4 benchmark. The taxonomy expands beyond conventional segmentation datasets by including classes such as autorickshaw, animal, and drivable surface distinctions such as drivable_fallback and non-drivable_fallback (Varma et al., 2018). The label system is organized as a four-level hierarchy: level 1 has 7 super-classes, level 2 has 16 mid-classes, level 3 has 26 groups of related fine classes, and level 4 has 30 fully distinct semantic classes. This hierarchy explicitly supports variable task difficulty, from 7-way to 30-way semantic segmentation.

The dataset statistics underscore the long-tail nature of the scene distribution. Road-side objects such as vegetation, buildings, and cars dominate the pixel counts, while rare but operationally salient classes such as animal, trailer, traffic_light, and bicycle fall below $10^5$ pixels. Challenging scenes include ambiguous road boundaries, dense crowds, multiple riders on one bike, unmarked drivable fallback zones, and heavy shadows or smog (Varma et al., 2018).

The benchmark results formalize the difficulty. Mean Intersection over Union is defined as

$$\mathrm{IoU}_{(c)} = \frac{\mathrm{TP}_{(c)}}{\mathrm{TP}_{(c)} + \mathrm{FP}_{(c)} + \mathrm{FN}_{(c)}}, \qquad \mathrm{mIoU} = \frac{1}{C}\sum_{c=1}^{C}\mathrm{IoU}_{(c)}.$$

On the hierarchy levels, DRN-D-38 achieves 85.9% at level 1, 72.6% at level 2, and 66.6% at level 3, while the best challenge entry, PSPNet ensemble, reaches 74.3% at level 3 (Varma et al., 2018). More consequentially, the cross-domain study shows a stark gap: CS→CS mIoU on common labels is 76%, ID→ID is 70%, CS→ID drops to 34%, and ID→CS is 49%. This 42-point fall from Cityscapes to IDD is one of the clearest empirical arguments for India-specific perception research.

3. Domain adaptation and boundless segmentation on Indian roads

The domain-gap problem is addressed directly by BoMuDANet, an unsupervised multi-source adaptation method for visual scene understanding in unstructured driving environments (Kothandaraman et al., 2020). The target domain is IDD, with source domains such as GTA5, Cityscapes, BDD, and SynScapes. The segmentation network $G(\cdot;\theta)$ is paired with supervised loss on a best source, self-training on target pseudo-labels, multi-source knowledge distillation, and an optional adversarial discriminator loss.

The overall objective is written as

$$\mathcal L = \alpha_{\rm sup}\,\mathcal L_{\rm sup} + \alpha_{\rm self}\,\mathcal L_{\rm self} + \alpha_{\rm distil}\,\mathcal L_{\rm distil} + \alpha_{\rm adv}\,\mathcal L_{\rm adv},$$

with typical weights $$\alpha_{\rm sup}=1,\alpha_{\rm self}=0.4,\alpha_{\rm distil}=0.5,\alpha_{\rm adv}=0.1$$ (Kothandaraman et al., 2020). The method also introduces the Alt-Inc self-training algorithm, an EM-style alternating-incremental loop in which pseudo-labels are iteratively updated and the network is refined end-to-end over roughly 3–5 rounds.

A distinctive contribution is “boundless” domain adaptation for open-set classes. For each pixel, the method computes the confidence

$s_{h,w} = \max_c P(h,w,c),$

and if the predicted class belongs to {car, truck, bus, motorbike} but $s_{h,w}<\tau$, with $\tau=0.85$ given as an example, the pixel is relabeled as open-set (Kothandaraman et al., 2020). This is intended to capture previously unseen classes such as autorickshaws without additional retraining.

On IDD, the best single-source baseline, AdvEnt with ResNet-101, yields mIoU $\approx 38.53\%$ and mAcc $\approx 86.68\%$. BoMuDANet with DRN-D-38 reaches overall mIoU $$\mathrm{IoU}_{(c)} = \frac{\mathrm{TP}_{(c)}}{\mathrm{TP}_{(c)} + \mathrm{FP}_{(c)} + \mathrm{FN}_{(c)}}, \qquad \mathrm{mIoU} = \frac{1}{C}\sum_{c=1}^{C}\mathrm{IoU}_{(c)}.$$0 and mAcc $$\mathrm{IoU}_{(c)} = \frac{\mathrm{TP}_{(c)}}{\mathrm{TP}_{(c)} + \mathrm{FP}_{(c)} + \mathrm{FN}_{(c)}}, \qquad \mathrm{mIoU} = \frac{1}{C}\sum_{c=1}^{C}\mathrm{IoU}_{(c)}.$$1, with excerpted per-class IoU values including Road 94.02%, Building 61.79%, Traffic-light 20.61%, Vegetation 81.75%, Car 94.16%, Truck 32.12%, Bus 4.67%, Motorbike 42.64%, and Bicycle 38.61% (Kothandaraman et al., 2020). The paper reports gains of +5.17% to +42.9% mIoU over prior state-of-the-art UDA and semi-supervised methods, along with real-time inference on IDD at approximately 2 fps using a 26.5 M-parameter model.

These results reinforce the notion that the central challenge is not only label scarcity but cross-domain brittleness under unstructured traffic, rural dirt roads, out-of-vocabulary categories, and severe class imbalance.

4. DriveIndia as a dataset family: 2D detection and 3D perception

The literature supports viewing DriveIndia as a dataset family for Indian traffic perception, with complementary 2D and 3D modalities.

Resource	Core content	Benchmark focus
IDD (Varma et al., 2018)	10,004 finely annotated images; 34 semantic classes; four-level hierarchy	Semantic segmentation in unconstrained environments
IDD-3D (Dokania et al., 2022)	15,500 total LiDAR frames; 223,000 3D boxes; six RGB cameras + one 64-beam Ouster OS1 LiDAR	3D object detection and 3D multi-object tracking
DriveIndia (Kumar et al., 26 Jul 2025)	66,986 RGB images; $$\mathrm{IoU}_{(c)} = \frac{\mathrm{TP}_{(c)}}{\mathrm{TP}_{(c)} + \mathrm{FP}_{(c)} + \mathrm{FN}_{(c)}}, \qquad \mathrm{mIoU} = \frac{1}{C}\sum_{c=1}^{C}\mathrm{IoU}_{(c)}.$$2 bounding boxes; 24 object categories	2D object detection with YOLO-family baselines

IDD-3D was collected over two 5-hour daytime sessions across varied Hyderabad neighborhoods, including flyovers, intersections, narrow streets, vendors, animals, and makeshift carts (Dokania et al., 2022). The sensor suite comprises one 64-beam Ouster OS1 LiDAR with 120 m radius and six synchronized high-resolution RGB cameras arranged for full 360° coverage. The released data include 12,000 LiDAR frames for train/val and 3,500 reserved for public test, with more than 5 hours of unlabeled raw rosbag streams for self-supervised and domain-adaptation research.

Its labels include 10 primary benchmark classes grouped into Vehicle, Rider, and Pedestrian super-categories, plus 7 additional miscellaneous classes such as Auto-rickshaw, HandCart, ConcreteMixer, Animal, and Tricycle (Dokania et al., 2022). More than 70% of all boxes lie within 25 m of the ego-vehicle, and the per-frame object density averages about 14 boxes within 30 m, with tails above 60 in crowded scenes. For 3D detection on the validation set over 10 classes and 0–30 m, CenterPoint without pretraining reaches mAP 49.99%, SECOND 44.31%, and PointPillars 39.27%. For tracking with SimpleTrack plus CenterPoint detections, overall AMOTA is 0.472 and AMOTP is 0.765, with Vehicle AMOTA 0.72 and Pedestrian AMOTA 0.28 (Dokania et al., 2022). Lower scores than nuScenes and KITTI are explicitly interpreted as evidence of higher complexity.

The 2025 object-detection dataset formally titled “DriveIndia” provides 66,986 high-resolution images at 1920×1080 resolution, annotated in YOLO format across 24 traffic-relevant object categories (Kumar et al., 26 Jul 2025). The class set spans pedestrians; vehicles such as Bicycle, Car, Motorcycle, Bus, Commercial vehicle, Truck, Auto-rickshaw, Ambulance, Police vehicle, Tractor, Pushcart, and Construction vehicle; road infrastructure such as Route board, Traffic sign, Traffic light, Temporary traffic barrier, Traffic cone, Rumble strips, Unmarked speed bump, Marked speed bump, and Zebra crossing; and anomaly classes Animal and Pothole. Data collection used vehicle-mounted front and back cameras over 120+ hours and approximately 3,400 km in southern India across urban roads, rural roads, and highways, including flyovers and construction zones (Kumar et al., 26 Jul 2025).

Its benchmark protocol uses

$$\mathrm{IoU}_{(c)} = \frac{\mathrm{TP}_{(c)}}{\mathrm{TP}_{(c)} + \mathrm{FP}_{(c)} + \mathrm{FN}_{(c)}}, \qquad \mathrm{mIoU} = \frac{1}{C}\sum_{c=1}^{C}\mathrm{IoU}_{(c)}.$$3

along with $$\mathrm{IoU}_{(c)} = \frac{\mathrm{TP}_{(c)}}{\mathrm{TP}_{(c)} + \mathrm{FP}_{(c)} + \mathrm{FN}_{(c)}}, \qquad \mathrm{mIoU} = \frac{1}{C}\sum_{c=1}^{C}\mathrm{IoU}_{(c)}.$$4, precision, and recall. All YOLO-family baselines were trained with input size 640×640, batch size 16, learning rate 0.01, and 100 epochs. Reported $$\mathrm{IoU}_{(c)} = \frac{\mathrm{TP}_{(c)}}{\mathrm{TP}_{(c)} + \mathrm{FP}_{(c)} + \mathrm{FN}_{(c)}}, \qquad \mathrm{mIoU} = \frac{1}{C}\sum_{c=1}^{C}\mathrm{IoU}_{(c)}.$$5 values are 74.5 for YOLOv5, 78.7 for YOLOv8, 77.2 for YOLOv9, and 79.0 for YOLOv11 (Kumar et al., 26 Jul 2025). The paper also identifies failure modes: mixed heterogeneous traffic and frequent occlusions hurt small and rare classes, with pushcart $$\mathrm{IoU}_{(c)} = \frac{\mathrm{TP}_{(c)}}{\mathrm{TP}_{(c)} + \mathrm{FP}_{(c)} + \mathrm{FN}_{(c)}}, \qquad \mathrm{mIoU} = \frac{1}{C}\sum_{c=1}^{C}\mathrm{IoU}_{(c)}.$$6 and unmarked speed-bump $$\mathrm{IoU}_{(c)} = \frac{\mathrm{TP}_{(c)}}{\mathrm{TP}_{(c)} + \mathrm{FP}_{(c)} + \mathrm{FN}_{(c)}}, \qquad \mathrm{mIoU} = \frac{1}{C}\sum_{c=1}^{C}\mathrm{IoU}_{(c)}.$$7; dense scenes and absent lane markings produce mis-localization of pothole and zebra crossing; and emergency vehicles remain long-tail classes.

5. Ride-hailing analytics and user segmentation

In mobility services, DriveIndia appears as a deployment-oriented interpretation of ride-hailing adoption analytics from Kolkata (Bhaduri et al., 2022). The study uses a three-step workflow: exploratory factor analysis identifies six latent attitude constructs; latent class cluster analysis segments respondents using those factor scores, with socio-demographics, land-use, and travel-habit variables entered as passive covariates; and three MCDM methods, MOORA, TOPSIS, and VIKOR, rank twelve ride-hailing-service attributes within each latent class, after which RankAggreg fuses the rankings into a single meta-ranking (Bhaduri et al., 2022).

The LCCA posterior class probability is given by

$$\mathrm{IoU}_{(c)} = \frac{\mathrm{TP}_{(c)}}{\mathrm{TP}_{(c)} + \mathrm{FP}_{(c)} + \mathrm{FN}_{(c)}}, \qquad \mathrm{mIoU} = \frac{1}{C}\sum_{c=1}^{C}\mathrm{IoU}_{(c)}.$$8

with a final solution of $$\mathrm{IoU}_{(c)} = \frac{\mathrm{TP}_{(c)}}{\mathrm{TP}_{(c)} + \mathrm{FP}_{(c)} + \mathrm{FN}_{(c)}}, \qquad \mathrm{mIoU} = \frac{1}{C}\sum_{c=1}^{C}\mathrm{IoU}_{(c)}.$$9 classes. The MCDM stage uses a normalization of the form

$G(\cdot;\theta)$0

with cost criteria inverted by $G(\cdot;\theta)$1 (Bhaduri et al., 2022).

The survey was conducted between March 1–20, 2021 across Kolkata’s 141 wards using a computer-aided, face-to-face household survey. After pilot-testing on 50 respondents, 1,000 adults aged at least 18 were targeted and 839 complete records remained after cleaning. The four questionnaire modules covered past-month RHS use, seventeen seven-point Likert statements spanning six latent attitudes, individual demographics, and household socio-demographics (Bhaduri et al., 2022).

Three latent user segments emerge. “Tech-savvy, Ride-Hailing-Ready” accounts for 48%, is strongest on tech-savviness and variety-seeking, weak on pro-environment and transit favourability, and shows the highest frequency of both commute and discretionary RHS use. “Traditional Active-Travellers” accounts for 28%, is high on favourability toward public transit and personal vehicles, lowest on tech-savviness, variety-seeking, and subjective norms, and 61% have never used RHS. “Multimodal, PV-Loving Individuals” accounts for 24%, is strong on vehicle favourability and environmental lifestyle, and uses RHS mainly for occasional discretionary trips (Bhaduri et al., 2022).

The meta-ranking of service attributes is stable across clusters. Motivators are ordered as Flexibility, Travel Time, Reliability, Availability, Safety, and Low Health Risk; deterrents are ordered as Travel Cost, Waiting Time, Driver Behaviour, Online Payment Issues, Customer Support, and App Interface (Bhaduri et al., 2022). Subtle differences remain: cluster 1 places Reliability above Travel Time; cluster 2 rates Safety above Availability; and cluster 3 shows the strongest sensitivity to Travel Cost. The proposed DriveIndia translation layer is explicitly segment-specific, including reliability guarantees and subscription bundles for the first cluster, voice or SMS booking and safety signaling for the second, and family-friendly shared shuttles plus cost-saving bundles for the third (Bhaduri et al., 2022).

6. Logistics 4.0, digital twins, and driver-assistance extensions

The logistics paper extends the DriveIndia label from traffic perception to industrial and supply-chain modernization (Shriyam et al., 2023). It describes India’s logistics costs as hovering near 15 percent of GDP and attributes inefficiency to fragmented infrastructure, siloed information flows, low automation, and the absence of real-time visibility. The proposed response combines Smart Industrial Parks, shared “Supply Hub in an Industrial Park” warehousing, IoT gateways, cloud-based monitoring, Digital Twins, optimization models, and policy instruments.

The CPS data flow is specified as

$G(\cdot;\theta)$2

with telemetry including availability, percent utilization, actual machining time, energy consumption, location, temperature, and vibration (Shriyam et al., 2023). DTs are divided into DTP, DTI, DTA, and DTE, with interoperability via open standards such as OPC UA and MQTT with TLS 1.3, authentication through OAuth 2.0 or X.509 certificates, and distributed learning supported by federated learning plus containerized microservices using Docker or Kubernetes (Shriyam et al., 2023).

The quantitative layer includes a SIP resource-allocation problem over shared resources $G(\cdot;\theta)$3 and jobs $G(\cdot;\theta)$4, with binary decision $G(\cdot;\theta)$5, objective

$G(\cdot;\theta)$6

subject to capacity and assignment constraints, and an end-to-end logistics cost minimization problem over suppliers $G(\cdot;\theta)$7, manufacturers $G(\cdot;\theta)$8, customers $G(\cdot;\theta)$9, flows $$\mathcal L = \alpha_{\rm sup}\,\mathcal L_{\rm sup} + \alpha_{\rm self}\,\mathcal L_{\rm self} + \alpha_{\rm distil}\,\mathcal L_{\rm distil} + \alpha_{\rm adv}\,\mathcal L_{\rm adv},$$0, $$\mathcal L = \alpha_{\rm sup}\,\mathcal L_{\rm sup} + \alpha_{\rm self}\,\mathcal L_{\rm self} + \alpha_{\rm distil}\,\mathcal L_{\rm distil} + \alpha_{\rm adv}\,\mathcal L_{\rm adv},$$1, and inventories $$\mathcal L = \alpha_{\rm sup}\,\mathcal L_{\rm sup} + \alpha_{\rm self}\,\mathcal L_{\rm self} + \alpha_{\rm distil}\,\mathcal L_{\rm distil} + \alpha_{\rm adv}\,\mathcal L_{\rm adv},$$2, $$\mathcal L = \alpha_{\rm sup}\,\mathcal L_{\rm sup} + \alpha_{\rm self}\,\mathcal L_{\rm self} + \alpha_{\rm distil}\,\mathcal L_{\rm distil} + \alpha_{\rm adv}\,\mathcal L_{\rm adv},$$3 (Shriyam et al., 2023). The paper also lists policy mechanisms including accelerated depreciation, tax credits, low-interest loans, R&D grants, standardization, skill-upgradation, sandboxes for blockchain-DT pilots, and integration with the National Logistics Policy.

At the road-infrastructure level, the pothole-oriented “Vehicle Driving Assistant” provides a narrower but operationally relevant extension (Dwivedi et al., 2020). Its pipeline uses video capture, frame extraction every 0.2 s, Gaussian smoothing, RGB-to-HSV conversion, road-color normalization, morphological cleaning, convex-hull road isolation, feature extraction, and classification. The top-performing model is Random Forest on 768-dimensional color-histogram features with $$\mathcal L = \alpha_{\rm sup}\,\mathcal L_{\rm sup} + \alpha_{\rm self}\,\mathcal L_{\rm self} + \alpha_{\rm distil}\,\mathcal L_{\rm distil} + \alpha_{\rm adv}\,\mathcal L_{\rm adv},$$4 and $$\mathcal L = \alpha_{\rm sup}\,\mathcal L_{\rm sup} + \alpha_{\rm self}\,\mathcal L_{\rm self} + \alpha_{\rm distil}\,\mathcal L_{\rm distil} + \alpha_{\rm adv}\,\mathcal L_{\rm adv},$$5, achieving 95.5% classification accuracy on the held-out test set; the reported example detection metrics are precision 0.94, recall 0.96, $$\mathcal L = \alpha_{\rm sup}\,\mathcal L_{\rm sup} + \alpha_{\rm self}\,\mathcal L_{\rm self} + \alpha_{\rm distil}\,\mathcal L_{\rm distil} + \alpha_{\rm adv}\,\mathcal L_{\rm adv},$$6-score 0.95, and box mIoU 0.82 (Dwivedi et al., 2020). In field trials on Bangalore city roads, the RF-histogram pipeline achieved a 92% detection rate at 15 m range with false alarm rate below 5% (Dwivedi et al., 2020). A plausible implication is that the inclusion of Pothole as a class in the 2025 DriveIndia object-detection dataset creates a direct bridge between classical driver-assistance pipelines and broader detection benchmarks (Kumar et al., 26 Jul 2025, Dwivedi et al., 2020).

7. Research significance and open directions

Across these works, DriveIndia is characterized by three recurring technical themes. First, Indian mobility environments require explicit modeling of heterogeneity: autorickshaws, animals, pushcarts, mixed two-wheeler traffic, unmarked or fallback drivable regions, emergency vehicles, potholes, and roadside anomalies appear as dataset classes or open-set adaptation targets (Varma et al., 2018, Kumar et al., 26 Jul 2025). Second, domain shift is structural rather than incidental, as shown by the Cityscapes-to-IDD degradation and the need for multi-source UDA, self-training, and open-set handling (Varma et al., 2018, Kothandaraman et al., 2020). Third, the research program extends beyond perception into service design and infrastructure orchestration, linking road-scene intelligence with ride-hailing adoption models and Logistics 4.0 architectures (Bhaduri et al., 2022, Shriyam et al., 2023).

Several future directions are already embedded in the cited works. IDD proposes domain adaptation, few-shot learning, behaviour prediction, robustification against dust and smog, and multi-task or curriculum learning over the four-level hierarchy (Varma et al., 2018). IDD-3D recommends class balancing, random rotations, scaling, LiDAR beam dropouts, domain-adaptive pretraining, synthetic insertion of rickshaws and animals, and future benchmarks for 3D semantic segmentation of vendor stalls and road-side objects (Dokania et al., 2022). The 2025 DriveIndia dataset recommends class-aware augmentation, multi-scale training, test-time augmentation, domain adaptation under fog and rain, self-supervised pre-training, and multimodal extensions using radar or LiDAR (Kumar et al., 26 Jul 2025).

The aggregate record therefore presents DriveIndia not as a single benchmark or product, but as a technically coherent research agenda centered on Indian mobility conditions. Its unifying contribution is the systematic elevation of local traffic structure, infrastructure irregularity, and operational heterogeneity into benchmark definitions, optimization models, and deployment workflows.