Book chapter
A Comprehensive Analysis of Different Object Detection Frameworks and Path Optimization Algorithms for an RGB-D Camera-Based Rock Detection System
Transforming Technologies in Smart Agriculture, pp 147-193
02 Aug 2024
Abstract
Artificial intelligence (AI) based decision-making for farm management, such as automatic rock detection and picking (field clearing) is gaining more attention. There hasn’t been much study done on a comprehensive comparison of state-of-the-art deep learning-based object detection architectures and their deployment to field rock detection systems. This study examined the different deep learning-based algorithms (e.g., Faster R-CNN, YOLOv3, YOLOv5 small, and EfficientDet) on rock detection. For each deep learning framework, we also provide a comprehensive examination of model design, training parameters, and testing on two sets of test datasets, of which the second dataset include previously unseen images for model assessment. To improve the robustness of the trained models, we perform data augmentation to increase the number of training images and induce diversity by performing various image transformation techniques during augmentation. In this study we also designed and built an RGB-D (red, green, blue, and depth) camera-based novel way to classify rocks and a moving arm to emulate the rock picking mechanism in order to achieve this aim. A prototype for the proposed system was built and tested in an indoor environment. In addition, we used optimization process to collect several rocks in the lowest cost. To achieve this, several path optimization algorithms were proposed and compared to decide the most efficient approximation heuristics that can perform optimization on a given graph in near-real-time. We also compare the performance of suggested object detection models before and after self-training on augmented data. Our findings indicate that the self-trained YOLOv5 small outperforms other evaluated models with an mAP@0.5 of 0.96, AP@0.5:0.95 of 0.69 and an F1-Score of 0.85. By comparing exact solvers and approximation algorithms such as: Local Search; Dynamic programming; Nearest Neighbor heuristics; Greedy heuristics; and Simulated Annealing for path optimization where the problem is treated as Traveling Salesman Problem (TSP) (NP-hard problem), our analysis rank Nearest Neighbor heuristic as the most efficient (0.0159 s for 1000 nodes) solver for achieving near real-time solution for optimized rock picking. The present research methods make a significant contribution to precision agriculture and can be adapted and applied to rock picking.
Metrics
7 Record Views
Details
- Title
- A Comprehensive Analysis of Different Object Detection Frameworks and Path Optimization Algorithms for an RGB-D Camera-Based Rock Detection System
- Creators
- Jithin Jose MathewPaulo FloresAnup Kumar DasYongxin JiangZhao Zhang
- Contributors
- Zhao Zhang (Editor)Yongxin Jiang (Editor)Chunhui Wen (Editor)Shuyuan Men (Editor)Yuan Zhang (Editor)
- Publication Details
- Transforming Technologies in Smart Agriculture, pp 147-193
- Series
- Smart Agriculture
- Publisher
- Springer Nature Singapore; Singapore
- Resource Type
- Book chapter
- Language
- English
- Academic Unit
- Electrical and Computer Engineering; College of Engineering
- Scopus ID
- 2-s2.0-85219720791
- Other Identifier
- 991021897315204721