In response to the rising adoption of smart video surveillance systems (SVS), this paper addresses a common challenge: occlusion in object detection. Existing object recognition methods often overlook the relative occlusion of neighboring objects, leading to real-world SVS systems encountering significant issues. Recent occlusion-handling approaches have limitations, such as data type inflexibility and difficulty distinguishing objects. In this study, we propose an end-to-end solution. Our approach utilizes the Feature Pyramid Network (FPN) for small and overlapping object detection, replaces Grey Level Co-Occurrence Matrices (GLCM) with point cloud density analysis for accurate occlusion rate determination, and integrates depth information with RGB data to improve occluded object separation. This holistic method aims to enhance object detection performance, particularly in challenging occlusion scenarios.