Construction Object Detection
Detecting construction resources (e.g., machines) in images or videos is the first and fundamental step required in the development of automation to analyze construction videos. Once construction objects have been correctly recognized, a large number of construction monitoring tasks could be automated. For example, detecting excavators and dump trucks at the same time could automatically calculate the dirt-loading cycles in earthmoving projects. The continuous detection of machines and workers can prevent potential collisions and alert construction engineers in a timely manner. Detection of construction materials identifies the material location in the supply chain, and enables effortless derivation of project performance indicators.
Annotations
Annotation Example
.jpg)
.jpg)
Example 1: The objects in the image that need to be annotated are A~E, F should not be annotated
Example 2: The objects in the image that need to be annotated are H, K and L, objects I and J should not be annotated
Annotation Standard
To ensure the annotation results are high quality, three standards a have been employed and were strictly followed during the object detection annotation process:
-
Consistency: all annotations in the proposed image data set are consistent in terms of the class definition, bounding box placement, and how to deal with occluded objects. The consistency also applies to how to deal with illuminations, how to control annotation quality, and how to annotate under snowing, raining, and night conditions.
-
Correctness: all annotations in the proposed image data set precisely describe the pixel-axis position of construction machine objects. The bounding boxes do not cut the objects, and the margins are within five pixels. The main objective is to annotate with as few errors as possible.
-
Completion: all objects belonging to the 10 classes are labeled. There may be more than one machine object in one image, and the annotations are exhaustive. All objects that can be identified were annotated, and occluded objects were annotated according to the part of the machine that can be seen.
Object Detection Algorithm Analysis
We have divided the ACID dataset into a training set (80%) and a validation set (20%). Four deep learning object detection algorithms have been tested on the ACID dataset, including Yolo-V3, Inception-SSD, Faster-rcnn-Resnet101, and R-Fcn-Resnet101.
​
Object Detection Demo Videos
The following video shows object detection of construction machinery with the model trained on the ACID dataset.