Sistema adaptativo de identificação e rastreamento para classificação e localização de objetos em tempo real baseado em câmeras para veículos autônomos

Abstract

The achievements of recent computer vision proposals show a proportional gain of object classification and localization accuracy to the complexity of neural network structures. The continuous focus on precision demands balance in processing to expand the number of applications and uses of a given system and guarantee its versatility. For this purpose, two computer vision approaches for object localization, detection and tracking, are described and associated. With current computer vision approaches and convolutional neural networks considered, a multistage algorithm is implemented adopting the YOLOv3 neural model for object detection and the HART approach for object tracking. In addition, we propose a post-detection adaptable stage change system based on the number of detected objects in sequence images and stage-related processing time. The adaptable system composes a model capable of presenting optimal performance results with detection and tracking features. In the end, we present a comparative performance evaluation, using the KITTI dataset, with the intention of analyzing and emphasizing the performance of each stage for sequences related to traffic environments. The evaluation, consisting of average precision and processing time metrics and definitions, is performed for multistage configurations of localization and post-localization tracking with adaptive stage changes.