Research On Driver Seat Belt Detection Algorithm Based On Semantic Segmentation

Seat belt detection of automobile drivers based on road monitoring is an indispensable part of the construction of intelligent transportation,and it is the unshirkable responsibility of traffic management departments to advise drivers to develop good driving habits of wearing seat belts and ensure the personal safety of drivers.With the development of artificial intelligence technology,detection methods based on machine learning and deep learning are widely applied to driver seat belt detection tasks,which greatly improves the efficiency of detection.However,the existing seatbelt detection algorithms based on machine learning have too high requirements for image quality,and the setting of the threshold of the algorithm is greatly disturbed by subjective interference and lacks robustness.The seatbelt detection algorithm based on deep learning mainly adopts a step-by-step strategy,which cannot perform end-to-end training detection,and the feature misalignment problem caused by ordinary horizontal detection frame to detect tilted seatbelts seriously limits the application of the model.The emergence of the new technology of rotated object detection algorithm has promoted the research of seatbelt detection algorithm based on rotated object detection.However,the seatbelt detection target in the actual detection scene is small,the shape is irregular,and the image quality is not high,which brings great challenges to the seatbelt detection of automobile drivers.(1)Aiming at the problems of small target and slow detection speed,a light and efficient rotating frame seat belt detection algorithm based on attention mechanism is proposed.Firstly,an attention-based feature fusion module(Attention Feature Fusion Module,AFFM)is constructed to integrate the features between different scales to enhance the feature expression of the target area;then align the tilt safety belt area to improve the feature malalignment;Finally,Mobilenet V2 is used as the backbone network to extract the features,effectively reduce the parameters of the model,reduce the computing power consumption and increase the detection speed of the model.The experimental results show that the average accuracy(AP)of the algorithm reaches 0.905,the recall(Recall)reaches 0.949,the number of parameters(Params)only needs 18.54 MB,and the end-to-end detection inference speed(FPS)reaches 14.6 images per second.Compared with some existing rotating frame target detection methods,it has significant advantages and can effectively improve the speed and accuracy of safety belt detection.(2)Aiming at the problems of low image quality,small target and difficult feature extraction in safety belt detection,the lightweight rotating frame belt detection algorithm based on attention mechanism is further explored,and a real-time rotating frame belt detection algorithm based on edge enhancement is proposed.Firstly,based on the PAFPN feature fusion network,Sharr edge detection operator is introduced to build edge enhancement attention feature network(Edge Enhance-attention Feature Fusion Module,EE-AFFM)to effectively enhance belt edge features,and use image enhancement means such as sharpening noise reduction to improve the problem of difficult to extract small belt features;then,introduce a one-stage real-time target detection network RTMDet based on YOLOX to improve the detection and reasoning speed.The experimental results show that the parameters of the algorithm needs 24.67 M,the computing power needs 99.76 GFLOPs,and the detection and reasoning speed is as high as 172 pictures per second.While ensuring the detection accuracy,the algorithm greatly improves the detection and reasoning speed,and meets the requirements of real-time industrial detection,and has a certain competitiveness.