Pedestrain Occlusion Detection Based On Deep Learning

Pedestrian detection has a wide range of applications,which can be applied in the fields of driverless,Video surveillance and security area and so on.It mainly refers to the process of classifying and locating the target in the video or picture.In the actual scene,due to the diversity of clothing for each person,the different levels of pedestrian occlusion,camera shooting angle and other human factors will lead to the bad result in existing pedestrian detection algorithms.At the same time,in complex scenes,there are natural factors such as illumination and weather,which are more difficult to the existing pedestrian detection algorithms.Therefore,nature factors and human factors bring great challenges to the existing pedestrian detection algorithms.In these challenges,the pedestrian occlusion detection is difficult to reduce the miss rate.So this paper is mainly reduce the miss detection because of the pedestrian occlusion.In this paper,we firstly analyze the research status of pedestrian detection algorithms in our country and abroad from traditional machine learning methods to deep learning methods in recent years.and then analyzes the reasons why the current pedestrian detection algorithms perform bad in the occlusion scene.Finally we propose two new occlusion pedestrian detection algorithms: RefinePedNet and Double Head RefinePedNet and evaluate the results in Citypersons dataset.The result shows that these two algorithms could alleviate the pedestrian occlusion problem.Furthermore,we also detect riders in the occlusion scene and our two detection algorithms could detect riders.In this paper,the main research work has the following 3 points.1)Data pre-process.Since the anchor based objective detection algorithms perform bad in the occlusion scene,inspired by CenterNet objective detection algorithm,we use anchor free algorithm to train the pedestrian data.Detailly,we change the input data from(x1,y1,x2,y2)to the center point and scale.In my experience,this method could alleviate the pedestrian occlusion problem,at the same time,it could also reduce the time consuming when the sliding window extract the anchor information.Furthermore,we split the human into 4 part: left body,right body,left leg,right leg.And then randomly select one part with a probability of 0.5 to ensure randomness.In my experience,this method could alleviate the pedestrian occlusion problem.2)A more efficient feature fusion method.In the real scene,Because of the difference between pedestrian and camera,the size of the target in the image is different,which will lead to miss detection.In this paper,we use multi-stage feature map fusion strategy,we combine the shallow feature map and the deep feature map and finally propose RefinePedNet pedestrian detection algorithm,thus it could alleviate the miss detection because the objects have different scale.3)Introducing double head mechanism.In my experience,RefinePedNet algorithm still has the bad result in severe occlusion,thus we add another branch in the detection head and fusion the detection result,finally it could alleviate the miss detection in the occlusion scene.In conclusion,pedestrian detection has a wide range of applications.However,it is inevitable to miss detection because of the pedestrian occlusion.Thus,in this paper,we mainly reduce the miss detection because of the pedestrian occlusion,at the same time,our algorithms perform well when detecting rider in the occlusion scene.We reduce the miss detection from three aspects: data pre-process,feature fusion,and introduce the double head mechanism.And these three method could reduce the miss detection.Compared with the state of art method,the two algorithms we proposed could reduce the miss rate in the heavy occlusion,which is achieved the research goal in our paper.