The rapid advancement of science and technology has led to promising research in computer vision, particularly in the field of object detection. Self-driving technology is a promising solution for preventing severe accidents. To achieve reliable autonomous driving, a deep understanding of the surrounding environment is crucial, often achieved through machine learning and deep learning, which convert sensory data into meaningful information. Object detection is a fundamental perception function of these systems and has been tackled through various approaches. In recent years, most proposed object detection frameworks have been two-dimensional (2D). However, these 2D methods cannot provide depth information essential for driving tasks like path planning, motion prediction, and collision avoidance, limiting their real-world applicability. Recent research has focused on 3D object detection. Despite this, there is a lack of organized knowledge in this area. Our objective is to investigate and study 3D object recognition methods that provide more accurate information about the object's size and location. In this thesis, we specifically concentrate on the deep learning-based approach. We categorize these approaches based on input data and focus on the point cloud-based approach. Our study introduces a novel method for 3D pedestrian detection in open environments using the DenseNet architecture and a spatial-semantic feature aggregation module called SSFA. Our method is applied to two advanced single-stage object detectors, CIA-SSD (Confidence IoU Aware Single-Stage Detector) and SE-SSD (Self-Ensembled Single-Stage Detector). By utilizing information from previous layers, the proposed approach enhances feature extraction, leading to improved pedestrian detection accuracy. We conclude by presenting and analyzing the proposed method and comparing its performance with previous techniques in terms of accuracy. The eva‎luation of the 3D pedestrian detection algorithm on the KITTI dataset demonstrates that our method achieves higher accuracy metrics, with a moderate average precision (moderate AP) of 44.03% on CIA-SSD and 44.49% on SE-SSD.

محمدرضا احمدزاده

نادر كريمي , حامد نريماني