Abstract:
Vehicle detection, which identifies vehicles as axis-aligned bounding boxes in still images, is widely used to estimate autonomous vehicle (AV) range, time-to-collision, and motion. Bounding boxes are convenient but too coarse to adapt to vehicle shape and pose variations. TBox (Trapezoid & Box) is a novel fine-grained representation for localization and recognition that extends the bounding box by restricting a vehicle’s spatial extent to a set of keypoints and indicating semantically significant local areas using subclasses. A cascaded anchor-free architecture estimates the bounding box and TBox, unlike monolithic models. One subnetwork detects vehicles as pairs of corners without anchors using a stacked hourglass network. It learns corner affinity fields for robust corner grouping. The other subnetwork estimates TBox keypoints. This subnetwork avoids ambiguous keypoint associations by reusing features. We propose a multitask learning strategy for training the cascaded model that implicitly integrates global context with local details, improving both tasks. A refinement algorithm explicitly corrects global box errors with robust local keypoints during testing, ensuring tight geometric representations for nearby critical vehicles. Our small model outperforms anchor-free detectors for vehicle detection and the TBox task.
Note: Please discuss with our team before submitting this abstract to the college. This Abstract or Synopsis varies based on student project requirements.
Did you like this final year project?
To download this project Code with thesis report and project training... Click Here
