Tuned YOLOv4 Model for Indoor 3D Object Detection from Point Cloud Data

3D object recognition using point cloud data is a popular field of study due to its numerous applications. Despite the large number of object detection models available in this field, the methods are unable to attain a higher accuracy rate. This is because most models are unable to adequately analyze the contextual information present in the data. Understanding the semantic meaning and the connection between the characteristics of different things is essential to producing accurate outcomes. The use of contextual information gleaned from the data to suggest a suitable response is the focus of this study. The point cloud data collected from the dataset is first spherically projected using the spherical coordinates. The primary semantic and geometric level information is then extracted by feeding the dual-level squeeze net architecture with the spherically projected images. These characteristics are in charge of giving the detection model context-level data so that it can produce precise results. After that, the collected features are sent into the tuned YOLOv4 model and the features are fused before inputting to the model. The proposed YOLOv4 model is giving best results with 96.65% accuracy.