Pedestrian Detection Research In Automated Driving Scene

Recent years the re-emergence of artificial intelligence lead the academic and the industrial to move toward new regulations.With artificial intelligence as the core,the new research route and theoretical system relying on deep learning technology have been gradually formed.The traditional algorithms,which are inefficient,complex,and the low universality has gradually,become a history.Throughout the entire industry,academia and research,there is no exception to the front line in the field.Driven by deep learning,computer vision has developed rapidly and achieved a series of results.The visual tasks such as target classification,target detection,semantic segmentation and instance segmentation are stepping through step by step,reaching an unprecedented height.Breakthroughs in technology are often following with scrambling commercial interests.Unmanned research companies such as Google,General Motors,and Baidu have invested heavily in unmanned research.As advanced driver-assisted core tasks,the pedestrian object detection based on deep learning comes into being.In the automatic driving scene,pedestrian objects often have different characteristics,such as different sizes,rich colors,widely distributed.In addition,the effects of imaging conditions,weather conditions,halo shadows and so on are not negligible,which are undoubtedly greatly increased the difficulty of research on this topic.Relies on the current deep learning technology,we'll design the most efficient model to obtain the most perfect detection effect,and to solve the problems of wrong detection and difficult detection about th small pedestrian target detection,and the missing detection and repeated detection in nighttime and other complex scene and so on.Based on these,the subject fully consulted the data before the experiment,investigated the data set and carried out detailed statistical analysis.In the experiment,the ability of the current pedestrian target detection model is excellent,such as spatial pyramid pooling,multi-scale feature fusion,depth separable convolution,residual connection,etc.,were applied to the model to design a very simple and efficient comprehensive detection model.During the training steps,the techniques of preheating training,multi-scale training and manual tuning are used.After training convergence,the channel pruning and weights compression techniques were applied to the model,which lays a foundation for the deployment of the model on the mobile side.The model obtained in the BDD 100 K dataset has a target mAP of 59%,while the FPS can maintain 70+,even for the the mAP of small pedestrian is also 38%.In order to further improve the detection of small targets,the experiment also tried to use croping the map,the region of interest extraction and infrared camera plus shielding visible light t,and which achieved good results.The experiment performed the channel pruning compression on the model,which reduced the model volume by 50% and the accuracy by only 2%.In addition,the accuracy of FP32 was converted to the precision of INT8 during the inference,which further increased the speed to 1.3 times,and the mAP was finally maintained.At 54.7%.By the research of this subject,not only the difficult problems in the current automatic driving field are well solved,but also various deep learning theories and the practicality of techniques are verified,which can better stimulate the research interests of researchers,and also can further promote the development of technology in the field.