基于自校正照明网络的苹果采摘机器人夜间目标识别方法

Target Recognition Method for Apple Picking Robot at Night Based on Self-correcting Illumination Network

摘要:

机器人夜间采摘是大幅提高苹果采摘机器人作业效率的有效途径。为解决夜间环境下苹果采摘机器人无法准确识别苹果，不能区分不同目标等关键问题，本文提出了一种基于自校正照明网络的苹果采摘机器人夜间目标识别方法。基于TracePro软件分析了图像采集区域的光照均匀性，确定了补偿光源最佳安装角度；对比分析白炽灯、LED灯、高压钠灯、卤素灯4种补偿光源下图像RGB、HSV分量的变化规律，确定白炽灯为适合苹果目标夜间识别检测的最优补偿光源；选用YOLO v8s作为主干网络，并将自校正照明网络SCINet嵌入到YOLO v8网络中，解决了图像局部过曝光和局部欠曝光问题；将注意力机制CBAM嵌入到骨干网络中，更好地提取苹果目标特征；用超轻量动态上采样器DySample模块替换原有的UpSample模块，增强对树冠深处小目标苹果的检测能力；在构建的夜间苹果数据集上进行训练与测试。试验结果表明，改进的YOLO v8模型可实现对夜间环境下图像中可采摘和不可采摘苹果的识别与分类，识别模型召回率为82.9%，准确率为81.4%，mAP为86.8%。将本文模型与YOLO v5s、YOLO v7、YOLO v8n、YOLO v8s、YOLO v8m、YOLO v11s等主流模型进行对比，本文改进的YOLO v8s模型mAP分别提高3.9、13.1、12.9、6.2、2.7、5.7个百分点，召回率分别提高0.9、6.1、7.2、2.6、2.3、1.8个百分点。研究结果可为苹果采摘机器人在夜间环境下采摘作业提供技术支撑。

Abstract:

Nighttime picking by robots is an effective way to significantly improve the operational efficiency of apple-picking robots. To solve the key problems such as the inability of apple-picking robots to accurately identify apples and distinguish different targets in the night environment, a night target recognition method for apple-picking robots was proposed based on a self-correcting illumination network. Based on the TracePro software, the illumination uniformity of the image acquisition area was analyzed, and the optimal installation angle of the compensation light source was determined. By comparing and analyzing the variation laws of the RGB and HSV component values of images under four compensation light sources, namely incandescent lamps, LED lamps, high-pressure sodium lamps, and halogen lamps, it was determined that incandescent lamps were the optimal compensation light sources suitable for the night recognition and detection of apple targets. YOLO v8s was selected as the backbone network, and the self-correcting illumination network SCINet was embedded into the YOLO v8 network, which solved the problems of local overexposure and local underexposure of the image. The attention mechanism CBAM was embeded into the backbone network to better extract the features of apple targets;the original UpSample module was replaced with the ultra-lightweight dynamic upsampler DySample module to enhance the detection ability of small targets such as apples deep in the tree canopy. Training and testing were conducted on the constructed nighttime apple dataset. The test results showed that the improved YOLO v8 model can achieve the recognition and classification of picking and non-picking apples in images under the night environment. The recall rate of the recognition model was 82.9%, the accuracy rate was 81.4%, and the mAP was 86.8%. The model was compared with mainstream models such as YOLO v5s, YOLO v7, YOLO v8n, YOLO v8s, YOLO v8m, and YOLO v11s. The mAP of the improved YOLO v8s model was increased by 3.9, 13.1, 12.9, 6.2, 2.7 and 5.7 percentage points respectively. The recall rates were increased by 0.9, 6.1, 7.2, 2.6, 2.3 and 1.8 percentage points respectively. The research results can provide technical support for the apple-picking robot’s operations in the night environment.