PROJECT TITLE :

Learning TBox With a Cascaded Anchor-Free Network for Vehicle Detection

ABSTRACT:

Vehicle detection, which is the process of identifying vehicles as axis-aligned bounding boxes in still images, is utilized quite frequently in order to estimate the range, time-to-collision, and motion of Autonomous Vehicles (AVs). Although convenient, bounding boxes are too coarse to adapt very well to the many different poses and shapes that vehicles can assume. In this work, we present TBox (Trapezoid & Box), a novel fine-grained representation that is useful for both localization and recognition. TBox extends the bounding box by limiting the spatial extent of a vehicle to a set of keypoints and indicating semantically significant local areas using subclasses. TBox is a novel fine-grained representation that is useful for both localization and recognition. We propose a cascaded anchor-free architecture to estimate the bounding box and TBox, which is a departure from the monolithic models that have been used in the past. Without making use of anchors, one of the subnetworks can recognize each vehicle as a pair of corners by employing a stacked hourglass network. In particular, it acquires knowledge of corner affinity fields, which enables it to carry out reliable corner grouping. The TBox is approximated as a collection of keypoints by the other subnetwork. This subnetwork makes use of the results of the bounding box to eliminate the possibility of ambiguous keypoint associations. Additionally, it recycles pre-existing features in order to cut down on the number of parameters. In addition to this, we suggest a multitask learning strategy for the training of the cascaded model. This strategy implicitly integrates the global context with the local details, which results in improvements for both tasks. A refinement algorithm makes explicit use of robust local keypoints during testing in order to correct any possible global box errors. This helps to ensure that geometric representations of nearby critical vehicles are accurate. The results of the experiments demonstrate that our method outperforms other anchor-free detectors currently available for vehicle detection, and it achieves better performance on the TBox task despite employing a smaller model.