Research On Vehicle/Pedestrian Target Detection Method Based On YOLO Deep Neural Network

Vehicle and pedestrian detection technology plays an important role in autonomous driving,vehicle flow monitoring,illegal vehicle identification,and traffic incident detection.Real-time and accurate vehicle and pedestrian detection is the basis and premise for intelligent vehicles to complete path planning and decision-making control,which is helpful for driving behavior risk management and improving driving safety.Deep neural networks have gradually replaced traditional vision algorithms as the mainstream method for vehicle and pedestrian detection.However,in the face of complex road scenes such as occlusion,low light and dense small targets,its detection performance drops significantly.In addition,the memory usage of complex models is large and the inference speed is slow,making it difficult to meet the real-time deployment requirements of mobile terminals.Therefore,based on the YOLO series of deep neural network models,this paper conducts research on vehicle/pedestrian target detection in complex scenarios on the basis of ensuring the lightweight and real-time performance of the model.The specific contents and innovations are as follows:(1)An intersection vehicle detection method based on the improved YOLOv4 model and DeepSort is proposed.Using MobileNetV2 as the backbone network of the YOLOv4 detection algorithm to extract target features,construct a lightweight target detection algorithm YOLOv4MobileNetV2;for the problem of the large state prediction error of the Kalman filter algorithm in the DeepSort algorithm in complex driving scenarios,combined with the LSTM model The target tracking results are detected and fused;the speed of vehicles at the intersection is estimated according to the distance and time of multi-target tracking in the area of interest.This method improves the efficiency of target detection,improves the phenomenon of frequent switching of tracking object labels,and the accuracy rate of vehicle speed detection reaches 94.5%.(2)A nighttime pedestrian detection method based on data domain transfer and YOLOX algorithm is proposed.First,the loss function of YOLOX is optimized,and the original confidence prediction loss is replaced by the Focal Loss loss to improve the imbalance of positive and negative samples caused by the proportion of foreground and background;secondly,in order to improve the detection accuracy under dark light conditions at night,a training strategy based on data domain migration is designed,which combines the larger-scale daytime data set with the smaller-scale nighttime data set after dark processing,and then performs training and testing after low-light enhancement.The experimental results show that the AP50 of this method reaches 81.4%,the model size is 17.2MB,and the detection frames per second(FPS)reaches 80,which meets the needs of real-time detection.(3)A nighttime vehicle pedestrian detection method based on the improved YOLOX model is proposed.First,the YOLOX model structure is re-parameterized and lightweight,and a coordinate-based attention mechanism is introduced into the backbone network;second,a feature scale fusion detection bypass is added to the feature pyramid,and the loss function is designed by combining the use of The CIoU method for target positioning and the Varifocal Loss method for confidence prediction improve the feature extraction ability for dense small targets and line-of-sight occlusion.After the method is trained by the data domain transfer strategy,the mean mAP of the average detection accuracy of vehicles and pedestrian targets at night is increased by 5.9%to 82.4%,which provides effective technical support for the safe driving of autonomous driving at night.