Reproduce “Joint Perception and Prediction for Autonomous Driving: A Survey” on OpenResearch

About this paper

Lucas Dal'Col, Miguel Oliveira, Vítor Santos

Joint Perception and Prediction for Autonomous Driving surveys 55 methods published from 2018 to 2024 that combine scene understanding, tracking, and future-motion forecasting in shared models. It organizes the field by sensor representation, scene-context modeling, and output format, and compares bounding-box, pixel-wise, and occupancy-based approaches while identifying open challenges in uncertainty, multimodal sensing, interaction modeling, supervision, and evaluation.

Key insights

• BEV remains the dominant input representation, while multi-view cameras and multimodal fusion are expanding; radar is still relatively underused despite its velocity and adverse-weather advantages.
• Explicit interaction modeling with graphs or transformers is becoming more common, but implicit map and interaction representations remain prevalent because they are simpler to implement.
• Occupancy and pixel-wise representations provide denser, more class-agnostic scene descriptions than bounding boxes, but they incur higher computational costs and can lose object identity or instance consistency.

Summary by alphaXiv.