Research And Implementation Of Human Detection Technology In Complex Cabin Environment

Human detection is the foundation and research hotspots of computer vision tasks,and it is also the core of video security monitoring.It also has great application prospects in the field of automatic driving,advanced auxiliary robot,safety monitoring and humanmachine interaction,and has great commercial value.Fishery and shipping industry play a decisive role in building a powerful marine country in China.However,fishery has been recognized as the most dangerous industry,most of which are caused by human factors.At present,limited by the bandwidth of intercommunication between sea and land by Beidou,the system cannot monitor the realtime video flow back and transmission of the ship's cockpit,so the safety monitoring and early warning of ship cockpit is very important for the safety and fishery status of the whole sea by means of human detection.It can also reduce the loss of life and property of the maritime employees,such as the fishermen.In recent twenty years,domestic and foreign scholars have conducted in-depth research in the area of object detection and pedestrian detection.Whether based on traditional feature training or deep convolutional network algorithm,the average accuracy has been greatly improved.But most of these algorithms is aimed at the general object and pedestrians,but the ship cockpit which is narrow,the personnel posture is diverse,the light is changeable and the shelter phenomenon happens frequently under the small scope complex scene,the traditional detection method receives the angle of view,the light and the disturbance and so on many aspects of challenges.In this paper,we conduct indepth research on human detection technology in cockpit complex environment,and transplant relevant models and algorithms into the safety monitoring system of fishing vessels,providing assistance for ship safety monitoring and early warning.In this paper,we rebuild the model based on ship cockpit scene.The basic idea is based on the features of multi component of the traditional deformable part model method and YOLO algorithm based on convolutional neural network detection method on the scene personnel targets were studied and experiments,and introduces the basic function and the design of ship safety monitoring system.In the traditional object detection experiments,we uses deformable part model which is the best algorithm of traditional feature extraction,and we rebuild the model for the ship cockpit indoor environment,select the HOG(Histogram of Gradient)features which is descended dimension,collect the dataset of cockpit personnel in the ship combined with the standard PASCAL VOC dataset the type of person.We also construct multiple components deformable part model,train the features with support vector machine,and optimize the weights among the components in the context of the scene and the non-maximum suppression to overcome the partial occlusion.Finally,we use multi-component model and multi-scale sliding window to detect,and train the new model to transplant to the ship safety monitoring system,and use multi-thread optimization to increase the detecting accuracy and speed.Based on deep learning method,we introduce the deep learning target detection based on the principle,model and algorithm.Limited by the performance of the general industrial control computer on the ship,we selected the fast detecting algorithm based on YOLO algorithm.To solve the problem of false detection and robustness of the algorithm,we used the surveillance video of the cabin to regenerate the training set.We trained the model on the basis of the pre-train model,applied it to the human detection in the cockpit,and introduced a simple YOLO-Tiny model to analyze and discuss the effect and performance of the network.It can be seen from the experiment that the average accuracy of the improved YOLO model can reach more than 90%,which proves the feasibility of transplanting to safe driving system.Although this article has made some discussion and research on human detection technology of ship cockpit under the complex environment,but also clearly see many deficiencies: the YOLO algorithm by deep learning on the CPU performance limit does not meet the requirements of practice which needs to optimize performance,also need to conduct analysis on human behavior,so as to better risk security warning.These problems still remain to be studied by subsequent researchers.