Plant Detection via Temporal Tracking and Region Merging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine vision techniques for detecting plants in images face challenges such as reduced accuracy due to overlapping leaves, occlusions, color changes, and the difficulty in distinguishing small plants, leading to missed detections and mislabeling of non-plant objects, along with issues related to data quantity, memory requirements, and time efficiency in training and inference.
Innovation Solution
A system and method that utilize a combination of machine learning models for detecting and tracking plants in a sequence of images, where a detector generates detection regions and a tracker updates and retains the position of plants across images, leveraging temporal information to improve detection accuracy and handle stationary plants with growth, by combining detection and tracking regions based on their positions and probabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are used to detect plants in images, then detection capability is provided, but accuracy decreases when leaves overlap or are occluded
Solution Approach 1:
The system performs preliminary actions by initializing trackers with detection regions from previous frames before processing current frame detections. This allows the system to pre-establish tracking hypotheses that constrain and guide subsequent detection, improving accuracy in challenging conditions where leaves overlap or are occluded.
Solution Approach 2:
The system implements feedback by using tracker states from previous frames to influence current detection results. The trackers provide feedback about expected plant positions and appearances, which are combined with new detection regions to refine accuracy. This feedback loop allows the system to correct detection errors that occur when leaves overlap or are occluded.
2Reliability
If detection is performed on every image in a sequence, then detection coverage is improved, but computational time and resource usage increase
Solution Approach 1:
The system merges detection operations with tracking operations, combining the strengths of both approaches. By integrating trackers that maintain state information across frames with periodic detection operations, the system achieves comprehensive detection coverage while reducing the computational burden of running full detection on every single frame.
3Measurement precision
If more training data is used to improve model accuracy, then detection accuracy improves, but memory requirements and training time increase
Solution Approach 1:
The system changes parameters by using tracker states to provide temporal context and constraints that effectively enhance the information content of each training example. This allows the model to achieve higher accuracy without proportionally increasing the quantity of training data, as the temporal relationships captured by trackers provide additional discriminative information.
Data Source
AI summary
Systems and methods for detecting plants in a sequence of images are provided. A plant is predicted to be in a detection region in an image and the plant is tracked across multiple images. A tracker retains a memory of the plants past position and updates a tracking region for each subsequent image based on the memory and the new images, thus using temporal information to augment detection performance. The plant can be substantially stationary and exhibit growth between images. Tracking substantially stationary plants can improve detection of the plant between images relative to detection alone. The tracking region can be updated based on the substantially stationary position of the plant, for instance by combining the tracking region with further predictions of plant position in subsequent images. Combining can involve determining a union.