Research And Application Of Embedded Pedestrian Detection System Based On Deep Learning

Human being is one of the most important recognition targets in computer vision applications.The accurate identification of human targets is an important condition for machines to complete subsequent tasks or interact with humans.Therefore,pedestrian detection task is one of the most popular research field in computer vision,which has important applications in security surveillance,self-driving car and industrial manufacturing.With the rapid development of artificial intelligence technology and the tremendous improvement of computer computing performance in recent years,pedestrian detection tasks and deep learning techniques can combined to get better performance.Compared with traditional pedestrian detection technology,GPU-based deep learning pedestrian detection technology worked.Both accuracy and detection speed are greatly improved.At the same time,with the rapid development of edge computing technology based on embedded platform,the deployment of pedestrian detection model on embedded platform for real-time pedestrian detection has become a research hotspot.In this thesis,we will study and implement a pedestrian detection system based on deep learning method on embedded platform.Due to the limitation of computing resources,this thesis will conduct in-depth research on YOLOv3 model based on deep learning one-stage method.Many optimization methods are applied to construct the network,which will be suitable for embedded platform computing requirements.At the same time,the structure and parameters are further designed and adjusted for pedestrian detection tasks,and the model lightweighting method--MobileNet is introduced to reduce the number of parameters.We used open datasets and take related integration processing,so that the model converges in the richest data scene under the condition of limited parameters.The pedestrian detection model designed and trained for embedded system computing resources is tested which shows a good performance.In addition,for the NVIDIA Jetson TX2 embedded platform,this thesis used the TensorRT library to design pedestrian detection to construct an accelerate inference framework,which achieves a large speed increase with less loss of accuracy.The V4L2 driver and Gstreamer video codec components can accelerate the video encoding and decoding of the embedded side,reducing the latency of end-to-end processing between video input and output.In this thesis,a vehicle-assisted pedestrian detection system is constructed,which integrates the designed model and acceleration framework into an embedded platform and combines with the pedestrian detection scene in the vehicle-assisted driving platform.When the pedestrian appears in the dangerous area of the driving direction,the system can send out warning signals in time to help the vehicle-assisted driving platform to make subsequent protective measures.The pedestrian detection system can detect middle and short distance pedestrians quickly and accurately.The working frame rate is above 21 FPS and the delay time is less than 0.5 seconds.It can realize real-time pedestrian detection under low-speed driving conditions.