From HMMs to Deep Learning: A Decade of AV Occupancy Prediction
Autonomous vehicle occupancy prediction encompasses the methods by which self-driving systems forecast the future spatial states of surrounding agents — vehicles, pedestrians, and unseen obstacles — to enable safe, collision-free navigation. This analysis draws on 70+ patent records and research publications spanning 2017–2025 to map the core technical approaches, leading assignees, geographic concentrations, and emerging directions in this domain.
Publication dates in the retrieved dataset span from 2016 to 2025, revealing a field that has transitioned from rule-based to learning-based approaches over roughly a decade. The foundational phase (2016–2018) relied on Hidden Markov Models and shallow neural networks: the Chinese Academy of Sciences’ 2017 paper introduced an ontology-based scenario-adaptive HMM approach, while a 2017 University of Central Florida paper proposed a two-layer neural network for vehicle-to-vehicle maneuver prediction. These represent pre-deep-learning baselines that the field has since substantially surpassed.
The development cluster of 2019–2021 marks the industrialization of deep learning–based prediction. TuSimple filed its foundational proximate vehicle intention prediction patents in US and WO in December 2019. Uber ATG published landmark work on uncertainty-aware motion prediction in 2020. IMRA Europe filed its PCT application for multimodal egocentric future prediction in December 2021. Stanford University proposed the Double-Prong ConvLSTM architecture for spatiotemporal occupancy in 2021. By the maturation phase of 2022–2025, the emphasis shifted toward integrated perception-prediction-planning pipelines and formal performance evaluation frameworks — with NVIDIA filing its bi-directional LSTM trajectory prediction patent in EP as recently as August 2025, and Five AI’s game-tree–based AV planning patent becoming active in EP in November 2025.
AV occupancy prediction patent filings span from 2016 to 2025, with the field transitioning from rule-based Hidden Markov Models in the foundational phase (2016–2018) to deep learning–based architectures during the industrialization phase (2019–2021), and toward integrated perception-prediction-planning pipelines in the maturation phase (2022–2025).
Four Technical Clusters Defining the AV Prediction Patent Landscape
The AV occupancy prediction field organizes around three core representation paradigms — Occupancy Grid Maps, Trajectory-Based Prediction, and Egocentric Future Localization — plus a fourth architectural cluster that couples prediction directly with planning. Each cluster has distinct IP owners, research backing, and commercial deployment contexts.
Cluster 1: Occupancy Grid Map (OGM) Prediction
Occupancy grid map approaches discretize the environment into a spatial grid where each cell is assigned a predicted occupancy probability, often across multiple future time steps. Huawei Technologies’ 2023 EP-active patent discloses a kinodynamic-weighted OGM approach that generates filtered predicted OGMs across multiple future timestamps, feeding a single consolidated map to a trajectory generator. UATC (Uber ATG’s successor entity) advances this with its goal-directed occupancy prediction patent (2021, US), which maps candidate path cells to predicted occupancy over a defined time horizon using road network topology — directly addressing multi-modality without mode collapse. Stanford University’s Double-Prong ConvLSTM (2021) provides research grounding by separating static and dynamic prediction prongs and fusing them to maintain dynamic object fidelity at long horizons.
An OGM discretizes the environment into a spatial grid in which each cell carries a probability of future occupation. In AV prediction, kinodynamic weighting, static/dynamic object segmentation, and sensor fusion are applied to improve the accuracy of per-cell occupancy probability estimates across multiple future time steps.
Cluster 2: Trajectory and Intention Prediction via Deep Learning
Per-agent trajectory distributions are predicted using recurrent (LSTM, GRU), convolutional, graph neural, or transformer architectures, with outputs typically being multi-modal distributions over future waypoints or maneuver categories. TuSimple’s 2019 US-active patent extracts perception features, generates a proximate vehicle trajectory, and passes it through a trained intention prediction model to produce a predicted intention and trajectory for downstream subsystems. GM Cruise’s 2024 prediction layer training patent introduces a drivable-area distance metric to update prediction layers during training, ensuring predicted waypoints remain geometrically consistent with the road network. NVIDIA’s 2025 EP-active patent uses a bi-directional LSTM to encode spatial and state features per observed object, predicting lateral and longitudinal maneuvers to determine future locations for both navigation and simulation.
On the research side, Uber ATG’s 2020 uncertainty-aware deep convolutional approach produces raster-image inputs for actor motion forecasting with explicit uncertainty quantification. Mercedes-Benz’s CRAT-Pred (2022) applies crystal graph convolutional networks with multi-head self-attention for map-free social interaction modeling, according to the published research record at IEEE.
Explore the full patent records for trajectory prediction and OGM approaches in PatSnap Eureka.
Explore AV Prediction Patents in PatSnap Eureka →Cluster 3: Egocentric and Multimodal Future Localization
From the ego vehicle’s camera — often monocular RGB — agents’ future positions are predicted in the image plane or bird’s-eye view, accounting for egomotion. IMRA Europe’s 2021 WO patent cascades a reachability prior network (RPN), reachability transfer network (RTN), and future localization network (FLN) to predict future locations and emergent objects in the ego vehicle’s camera view, handling partial visibility and egomotion drift. Three counterpart filings (WO 2021, EP 2021, US 2023) give IMRA Europe broad multi-jurisdictional coverage for this framework. The University of Freiburg’s reachability prior method (2020) and Indiana University’s egocentric vehicle localization paper (2019) provide the academic foundation for this cluster.
Cluster 4: Planning-Integrated and Safety-Aware Prediction
Planning-integrated prediction architectures do not decouple prediction from planning: the ego vehicle’s candidate trajectories inform the prediction of surrounding agents’ behaviors, and contingency plans are generated simultaneously. Five AI’s 2025 EP-active patent executes tree-search through a game tree whose nodes represent anticipated driving scenario states, incorporating predicted external agent behavior at each node to produce safe maneuver sequences. TuSimple’s 2020 US-active patent couples ego vehicle trajectory planning with proximate vehicle trajectory prediction, checking for conflict before committing to a planned trajectory. GM Cruise’s 2024 yield prediction patent applies an ML-trained yield prediction model per entity to determine when the AV should yield, directly coupling entity trajectory prediction to AV path decisions.
“The isolated predict-then-plan pipeline is giving way to integrated architectures — Waabi’s LookOut, Five AI’s game tree, TuSimple’s conflict-checking planner — where prediction and strategic planning are tightly coupled at each decision node.”
Who Is Filing and Where: Assignee and Jurisdiction Analysis
Among 20+ distinct patent assignees in this dataset, five industrial players account for the majority of filing volume: UATC/Uber ATG, TuSimple, GM Cruise Holdings LLC, IMRA Europe, and NVIDIA. Their strategies differ materially in jurisdictional scope, filing cadence, and technical focus.
TuSimple holds the broadest jurisdictional coverage in this dataset, with 5 patent records across US, WO, and AU for proximate vehicle intention prediction and trajectory planning — demonstrating a deliberate multi-jurisdictional strategy that treats AU coverage as commercially relevant alongside the standard US/EP/WO cascade. UATC/Uber ATG holds 4 active US patents covering goal-directed occupancy prediction and performance metrics, complemented by a substantial research publication record; its dominance in OGM and performance evaluation reflects Uber ATG’s deployment to a production AV fleet, as described in the 2020 research paper. GM Cruise Holdings LLC filed 3 patent records — all in 2024 — covering prediction layer training, yield prediction, and intersection prediction, signaling recent acceleration rather than long-standing IP accumulation. IMRA Europe executed a textbook WO/EP/US cascade for its egocentric future prediction framework (WO 2021, EP 2021, US 2023), securing coverage across all major ADAS and AV deployment markets.
In the AV occupancy prediction patent dataset analyzed for this report, the US is the dominant filing jurisdiction with approximately 12 patent records, followed by EP with approximately 10 records and WO with 3 records. AU and JP each appear once, while China is represented through academic research but not through Chinese-jurisdiction patents in the retrieved sample.
The jurisdictional picture is notable for what is absent as much as what is present. China is represented through academic research from the Chinese Academy of Sciences, Xi’an Jiaotong University, Tsinghua University, Jiangsu University, Southwest University, and Dalian University of Technology — but not through Chinese-jurisdiction patents in this dataset. This is potentially a gap in the retrieved sample rather than a true absence of CN filings, and IP strategists monitoring Chinese assignees such as Baidu Apollo and Huawei should conduct targeted CN-jurisdiction searches to assess actual filing density. Korea appears through KAIST research and StradVision’s EP patent (Korean assignee). Japan appears through Toyota Motor Corporation’s JP-active parking availability patent. According to WIPO‘s global IP filing data, cross-jurisdictional AV patent activity has grown substantially since 2019, consistent with the pattern observed here.
Given the volume of Chinese university research on occupancy and trajectory prediction in this dataset, IP strategists should conduct targeted CN-jurisdiction searches to assess actual filing density from assignees such as Baidu Apollo and Huawei that are not captured in the retrieved sample.
Application Domains Beyond Highway Driving
AV occupancy prediction patents in this dataset span six distinct application domains, from on-highway trucking to parking lot management — each with different prediction horizon requirements, sensor configurations, and regulatory contexts.
The dominant application is autonomous road vehicles for on-highway and urban use. TuSimple’s entire patent portfolio targets commercial AV trucking and urban driving. GM Cruise’s prediction layer and yield prediction patents target urban robotaxi operations. NVIDIA’s trajectory prediction patent explicitly addresses both AV navigation and simulation. Uber ATG/UATC’s occupancy prediction work was deployed to a production AV fleet, as described in the 2020 research paper. Baidu’s Apollo platform paper (2020) describes a deployed prediction pipeline supporting geo-fenced AV operations across multiple cities.
Advanced Driver Assistance Systems (ADAS) represent the second major domain. IMRA Europe’s egocentric future prediction patents explicitly cover both fully autonomous vehicles and ADAS equipped with cameras, broadening the application to SAE Level 2–3 systems. Volvo Car Corporation’s 2019 EP-active patent applies prediction to determine when a vehicle can safely engage autonomous mode — a direct ADAS use case. The NHTSA and UNECE regulatory frameworks for ADAS increasingly require predictive safety evidence, making this application domain commercially significant for patent holders.
Fleet and multi-vehicle management is an emerging application. Aptiv Technologies’ 2024 EP-active patent uses predicted demand and event information to pre-position and task fleets of AVs. The University of Massachusetts’ 2024 EP-active patent models merge and obstruction scenarios with imminency factors for fleet-level operational control. V2X-augmented prediction also surfaces: StradVision’s 2024 EP patent integrates V2X data with image-based perception to predict interference from both V2X-capable and non-V2X vehicles, while Uber ATG’s V2VNet (2020) demonstrates joint perception and prediction via compressed feature sharing across nearby AVs.
UATC’s 2022 US patent and its 2025 continuation define avoidance and availability metrics that directly evaluate occupancy prediction model outputs against real and simulated trajectories — signaling that prediction model evaluation frameworks may become tied to regulatory certification, creating IP leverage for metric owners and compliance obligations for deployers.
Emerging Directions in 2023–2025 Filings
Five directions are gaining momentum based on the most recent filings in this dataset (2023–2025), each representing a shift from the established architectural assumptions of the 2019–2021 development cluster.
1. Formal performance metric frameworks for prediction models. UATC’s updated 2025 US-active patent defines complementary avoidance and availability metrics to evaluate ML object prediction models — signaling that the field is moving beyond accuracy benchmarks toward safety-relevant operational metrics. TuSimple’s 2023 US-active patent extends prediction to non-operational design domain (ODD) boundary detection, a critical safety envelope concern. Together, these filings suggest that prediction model evaluation frameworks may soon be tied to regulatory certification.
2. Game-theoretic and tree-search planning with integrated prediction. Five AI’s 2025 EP patent constructs game trees where each node’s anticipated state is determined by both AV maneuver candidates and predicted external agent behavior. This tight coupling of prediction and strategic planning represents a departure from decoupled predict-then-plan architectures and aligns with the academic direction set by Waabi’s LookOut system (2021), which optimizes contingency plans over a diverse joint distribution of multi-agent futures.
3. Attention-based interaction selection for scalable prediction. Mercedes-Benz’s pending 2025 US patent uses an attention-based interaction algorithm to identify only the most interaction-relevant vehicles for trajectory prediction — directly addressing the scalability challenge of multi-agent environments. This echoes the Motional 2022 research paper showing that attention weights from prediction models can rank agent importance.
4. V2X-augmented prediction for mixed-capability traffic. StradVision’s 2024 EP patent and Uber ATG’s V2VNet work (2020) both target environments with a mix of V2X-capable and legacy vehicles. The emergence of filed patents in this area — rather than just research papers — signals commercial readiness of V2X-integrated prediction pipelines.
5. Simulation-grade trajectory prediction. NVIDIA’s 2025 EP patent explicitly targets both AV navigation and simulation control of virtual objects using the same prediction architecture, pointing toward unified real-world/simulation prediction stacks that reduce the sim-to-real gap in AV testing pipelines. UC Berkeley’s Pylot platform (2021) provides a latency-accuracy evaluation framework for AV pipeline components including prediction, reinforcing the academic basis for this direction. For context on global AV simulation standards activity, the ISO TC22/SC32 working group has been developing standardized test scenario frameworks relevant to simulation-based prediction validation.
Map the full emerging directions landscape for AV prediction with PatSnap Eureka’s AI-native patent analysis tools.
Analyse AV Prediction Trends in PatSnap Eureka →Strategic Implications for IP and R&D Teams
The patent landscape for AV occupancy prediction presents four actionable strategic signals for IP counsel, R&D leaders, and product teams building or licensing prediction technology.
IP white space exists in safety-aware prediction for occluded agents. Academic work from Xi’an Jiaotong University (2021) addresses unseen vehicle occupancy maps — predicting earliest-occupancy maps for both seen and unseen vehicles — but among retrieved patents, no granted patent directly claims this capability. R&D teams should monitor filing activity in this sub-domain and consider prosecution strategies targeting occluded-agent prediction.
Jurisdictional diversification is a competitive priority. TuSimple’s simultaneous US/WO/AU filings and IMRA Europe’s WO/EP/US cascade demonstrate that leading players treat multi-jurisdictional coverage as essential. Single-jurisdiction filing strategies are likely insufficient for IP protection in this field, particularly given the global deployment ambitions of commercial AV operators. PatSnap’s patent analytics platform provides cross-jurisdictional filing gap analysis to support this kind of strategy.
Prediction-planning coupling is the emerging architectural frontier. The isolated predict-then-plan pipeline is giving way to integrated architectures. Product developers building prediction modules should design with planning-coupling APIs in mind, and IP teams should evaluate whether existing prediction claims can be extended to cover joint prediction-planning system claims.
Performance metrics are becoming standardizable and potentially regulatorily relevant. UATC’s avoidance/availability metric patents, combined with TuSimple’s ODD prediction system, suggest that prediction model evaluation frameworks may soon be tied to regulatory certification — creating IP leverage for metric owners and compliance obligations for deployers. Teams involved in AV certification should audit exposure to these metric patents as part of their freedom-to-operate analysis. The PatSnap IP management suite supports FTO workflows across multi-jurisdictional patent portfolios.
“IP white space exists in safety-aware prediction for occluded agents — academic work addresses unseen vehicle occupancy maps, but among retrieved patents, no granted patent directly claims this capability.”