Research On Vehicle Detection And Tracking Algorithm Based On Deep Learning

In recent years,with the rapid development of China's modernization process,the number of cars are increasing dramatically,leading to frequent traffic accidents at the same time.Major automakers are paying more and more attention to technologies for active and passive safety of automobiles and assisted or automatic driving.Object detection and tracking system is the basis of the environmental perceptual module,one of the important components for automatic driving system.Designing an object detection and tracking model with excellent performance is one of the research priorities and difficulties in the field of autonomous driving.Traditional artificial feature-based object detection algorithms have poor generalization ability and robustness.Therefore,this paper,based on deep learning,conducts a research on target detection and tracking algorithms for vehicles in road traffic and a vehicle detection and tracking model with both accuracy and real-time performance is proposed.Tasks have been done as follows:(1)SE-Tiny YOLOv3,an object detection algorithm based on deep learning,is proposed.Aiming at the problem that Tiny YOLOv3 detection model can not balance the accuracy and realtime well,depthwise separable convolutions are used to replace standard convolutional layers of Tiny network structure.The lateral depth of Tiny backbone is expanded based on SENet submodule,improving the feature learning ability of convolutional neural networks.Loss function for bounding box regression of YOLOv3 is modified to alleviate scale sensitivity.(2)Tricks of training deep learning models are studied.A simple classification model is built to analysis the effects of various optimization algorithms,learning rate decay schedulers and data augmentation methods on model accuracy on Fashion-MNIST and CIFA-10 classification datasets.Experiment results indicate that higher model classification accuracy can be achieved based on training tricks of SGDM optimization algorithm,cosine annealing learning rate scheduler and data augmentation skills by comparing the loss and accuracy curves during training,providing a preliminary basis for training detection model afterwards.(3)An object detection model containing 20 categories is trained on the VOC dataset using proposed SE-Tiny YOLOv3.The results show that SE-Tiny YOLOv3 with GIOU loss scores 62.6% mAP with a 9.9% boost over Tiny YOLO and 3.3% mAP higher than original loss function.The detection speed is 35 frames per second which can be used for real-time detection.Meanwhile,compared to Tiny YOLO,the model weight parameters decrease by 7.1MB to 28.3MB.(4)A car detection model is trained on KITTI and homemade dataset based on SE-Tiny YOLOv3 algorithm.Results show that SE-Tiny YOLOv3 with GIOU achieves a mAP of 90.0%,an increase of 8.7% before modifying the loss function,and the detection speed on a single image is 47 ms,a reduction of 11 ms.(5)A vehicle tracking algorithm is proposed.Tracking model is designed using Kalman filter algorithm.Both tracking and detection bounding boxes are correctly associated with Hungarian matching algorithm.On KITTI tracking sequence,the experimental results show that,compared to a merely detection algorithm,phenomena of false and missed detections are effectively suppressed after using detection model containing tracking algorithm.Our vehicle detection and tracking algorithm has a certain universality,which can provide a reference for the design of other object detection and tracking models.