YOLOv4-tiny网络夜间行人检测算法研究

Research on nighttime pedestrian detection algorithm for YOLOv4-tiny network

无人驾驶系统中行人检测对于交通安全起着至关重要的作用,由于夜间红绿蓝三原色(red,green,blue,RGB)图像色彩信息少、对比度差异大,对行人检测带来挑战性,为提高夜间场景下无人驾驶行人检测的准确性,提出了SCP-YOLOv4-tiny夜间行人检测算法。引入注意力机制,增强网络对夜间行人特征提取能力;添加空间金字塔池化模块,以丰富深度特征信息;采用SiLU激活函数替换YOLOv4-tiny算法原有激活函数,提高算法检测精度。在NightOwls与BDD100K公开数据集上对改进的算法进行训练与测试,结果表明,改进后的算法平均准确率达到了94.11%,较改进之前提高了16.84%,F F1-score值达到了0.92。使用无人驾驶实验平台采集夜间道路行人图像,并在车载硬件平台Jetson AGX Xavier使用SCP-YOLOv4-tiny算法对采集数据进行检测,验证了算法改进的有效性,能够满足无人驾驶系统应用需求。

Pedestrian detection in unmanned systems plays a crucial role for traffic safety.However,the low color information and large contrast differences of red,green,blue(RGB)images at night pose challenges for pedestrian detection.To improve the accuracy of unmanned pedestrian detection in nighttime scenarios,the SCP-YOLOv4-tiny nighttime pedestrian detection algorithm is proposed.The attention mechanism is introduced to enhance the network’s ability to extract nighttime pedestrian features;the spatial pyramid pooling module is added to enrich the depth feature information;the original activation function of the YOLOv4-tiny algorithm is replaced by the SiLU activation function to improve the detection accuracy of the algorithm.The improved algorithm is trained and tested on NightOwls with BDD100K public dataset,and the results show that the average accuracy of the improved algorithm reaches 94.11%,which is 16.84%higher than that before the improvement,and the F F1-score value reaches 0.92.The nighttime road pedestrian images are collected using a driverless experimental platform,and the in-vehicle hardware platform Jetson AGX Xavier uses SCP-YOLOv4-tiny algorithm to detect the collected data,which verifies the effectiveness of the algorithm improvement and can meet the requirements of unmanned system applications.