Research On The Deep Learning Based Hand Detection In Vehicles

As one of the important media for human to transmit information,human hand has been widely studied in the field of computer vision.In intelligent vehicle scenarios,hand detection in vehicles is not only the basis of Human Vehicle Interaction(HVI),but also the basis of analyzing driver's behavior and researching vehicle driving safety.To this end,this paper will study the deep learning based algorithms for hand detection in vehicles.Considering the accuracy and speed of the algorithm overall,We present a hand detection method named Multi-Scale YOLOv2 based on the efficient and fast one-stage object detection network YOLOv2.The proposed Multi-Scale YOLOv2 improves the performance of YOLOv2 on hand detection through three modules.First,Multi-Scale Feature Refinement Module is used to obtain more fine-grained features and improve the detection performance of the network for low-resolution hand.Second,Channel Importance Evaluation Module is introduced to automatically learn the importance of feature channels and re-weight the features to strengthen the representational ability of the features.Third,Hard Example Punishment Module utilizes a term of hard example loss to enhance the discriminant ability of the network,and eliminate more false detections.Aiming at the problem that the accuracy of hand detection algorithms based on the onestage object detection network is not high enough,this paper applies the two-stage object detection network Faster R-CNN to hand detection.ResNet101 is used as the basic feature extraction network to extract more expressive features.ROI Align is used instead of ROI Pooling to solve the problem that the position and size of the target candidate box do not match the groundtruth,improving the accuracy of hand detection.To solve the problem of hand deformation,the deformable convolution is introduced into the network to enhance the ability of deformation modeling and improve the performance of hand detection.Aiming at the problem of too slow speed and too large model of hand detection algorithms based on the two-stage object detection network,we propose Thin Faster R-CNN hand detection algorithm based on knowledge distillation.Firstly,the channels of the Faster R-CNN's feature extraction network ResNet101 is reduced to 1/4 of the original,and the Thin Faster R-CNN network is obtained.Then,the Thin Faster R-CNN is trained based on knowledge distillation,including attention-based featuremap knowledge distillation and classification knowledge distillation.Among them,the attention-based featuremap knowledge distillation encourages the student network to learn similar feature representations as the teacher network,the classification knowledge distillation enables students network to learn a more powerful classifier.Experiments show that the hand detection speed of Thin Faster R-CNN can reach real-time on GPU.After knowledge distillation training,the hand detection accuracy of Thin Faster R-CNN has been greatly improved.In this paper,the accuracy and speed of the algorithm are studied,which greatly improves the efficiency of hand detection in vehicle scenes.It is of great significance in the application of Human Vehicle Interaction(HVI)system and the research of vehicle driving safety.