FL-YOLO Multi-target Detection System Based On YOLOV3 Algorithm

In recent years,self-driving cars have created a wave of research in the automotive and artificial intelligence industries.In the automatic driving technology of automobiles,the recognition and discrimination of target objects of complex environments is a difficult challenge and one of the key tasks to be solved.Target detection is not only an important branch of computer vision research,but also a key link and task in autonomous driving systems.Due to the complexity of actual road conditions,it is difficult to greatly improve the technicality of assisted car driving based on traditional target detection.In the target detection application of automatic car driving,vehicle detection and pedestrian detection are the most common detection tasks.In the actual car driving road conditions,the detection effect of the middle vehicle and the pedestrian in the image of the front car of the automatic driving is easily affected by illumination,shooting angle,weather,occlusion and the like.At present,R-CNN series algorithms based on candidate boxes are the most commonly used deep learning algorithms in the field of target detection.Although this series of algorithms has high detection accuracy in the experiment,due to the complexity of the network,it will cause high delay in real-time detection,making it difficult to popularize this series of algorithms in industrial applications.The end-to-end YOLO series algorithm based on regression method not only reduces the complexity of convolutional network,but also meets the real-time detection requirements in industrial applications.Based on the latest YOLOv3 algorithm in the industry,this paper proposes a FL-YOLO system for multi-target detection of vehicles and pedestrians in an autonomous driving environment.In this paper,based on the YOLOv3 algorithm,there are problems with the multi-target detection of vehicles and pedestrians in automatic driving.The optimization of the FL-YOLO system is proposed to the following three aspects :(1)In order to solve the problem of missed detection of large-scale targets for road conditions images by the YOLOv3 algorithm during vehicle driving,the FL-YOLO system improves the feature extraction network structure of YOLOv3 and adjusts the multi-scale to obtain a larger field of view than YOLOv3 Feature map of size.(2)In order to solve the problem of low model recalls rate and missed detection caused by the mutual occlusion of pedestrians and vehicles in actual urban target dense road conditions,the FL-YOLO system borrowed the idea of repulsion loss function and(3)introduced the model training loss calculation Two loss exclusions.(4)The FL-YOLO system adopts a parallel processing mode on the architecture,and uses a buffer queue of YOLO detection functions to reduce the calculation waiting time for the detection to build a lightweight network model.