智能盘点钢筋数量AI识别

智能盘点—钢筋数量AI识别

赛题背景

在工地现场，对于进场的钢筋车，验收人员需要对车上的钢筋进行现场人工点根，确认数量后钢筋车才能完成进场卸货。目前现场采用人工计数的方式，如图1-1中所示：
（At the construction site, the acceptance personnel shall manually count the number of rebars loaded on the trucks entering the site, and the trucks can unload only after the quantity is confirmed. Manual counting is currently used at the site, as shown in Figure 1-1: ）

图1-1 钢筋点跟现场场景

上述过程繁琐、消耗人力且速度很慢（一般一车钢筋需要半小时，一次进场盘点需数个小时）。针对上述问题，希望通过手机拍照->目标检测计数->人工修改少量误检的方式（如图1-2）智能、高效的完成此任务：
（The process is tedious, labor-intensive and slow (usually it takes several hours to count a truckload of rebars). In order to solve the above problems, we hope to accomplish this task intelligently and efficiently through mobile phone photographing-> target detection and counting-> manual little errors modification (as shown in Figure 1-2): ）

图1-2 理想工作场景

主要难点：
（Main Difficulties:）
（1）精度要求高（High precision requirement ）
钢筋本身价格较昂贵，且在实际使用中数量很大，误检和漏检都需要人工在大量的标记点中找出，所以需要精度非常高才能保证验收人员的使用体验。需要专门针对此密集目标的检测算法进行优化，另外，还需要处理拍摄角度、光线不完全受控，钢筋存在长短不齐、可能存在遮挡等情况。
（The rebar is expensive and extensively used. Therefore, errors and omissions must be manually identified from a large number of marked points, and only high precision can ensure the user-experience of acceptance personnel. Accordingly, the detection algorithm for this dense target needs to be optimized. Furthermore, problems in relation to the shooting angle, incomplete control of light, uneven length of rebars, and possible occlusion must be addressed.）
（2）钢筋尺寸不一(Various dimensions of rebars)
钢筋的直径变化范围较大（12-32中间很多种类）且截面形状不规则、颜色不一，拍摄的角度、距离也不完全受控，这也导致传统算法在实际使用的过程中效果很难稳定。
(The diameter of rebars varying from 12 to 32 mm, the irregular section shapes, uniform colors and incomplete control of the shooting angle and distance have led to result instability of the traditional algorithm in actual use. )
（3）边界难以区分(Indistinguishable boundaries )
一辆钢筋车一次会运输很多捆钢筋（如图1-3），如果直接全部处理会存在边缘角度差、遮挡等问题效果不好，目前在用单捆处理+最后合计的流程，这样的处理过程就会需要对捆间进行分割或者对最终结果进行去重，难度较大。
(A truck will carry many bundles of rebars at one time (as shown in Figure 1-3). If all the rebars are counted directly, there will be problems such as edge angle difference and occlusion. The current method of single bundle counting + final summation needs to divide the bundles or remove the duplicates of the final results, but it is difficult. )

图1-3 钢筋进场场景

赛题任务

本赛题基于广联达公司提供的钢筋进场现场的图片和标注，希望参赛者综合运用计算机视觉和机器学习/深度学习等技术，实现拍照即可完成钢筋点根任务，大幅度提升建筑行业关键物料的进场效率和盘点准确性，将建筑工人从这项极其枯燥繁重的工作中解脱出来。
比赛任务以算法在验证数据集上的精度为主要评估指标，具体技术指标以《评分方式》一节中给出的量化指标为准。
(Based on the pictures and labels provided by Glodon for the entry of rebars, contestants are expected to combine computer vision and machine learning/deep learning techniques to accomplish the task of counting the number of rebars by taking photographs, thereby greatly improving the entry efficiency and inventory accuracy of key materials in the construction industry, and relieving construction workers from this extremely boring and onerous task.
The task of the competition is to evaluate the accuracy of the algorithm on the verification data set, and the specific technical indicators are subject to the quantitative indicators given in the section of Scoring Method. )

link

code 

link

link