Research On Lightweight Deep Learning Algorithms For Pedestrian Detection In Unmanned Driving

Pedestrian detection technology is one of the key technologies for driverless vehicles.Whether driverless cars can detect pedestrians quickly and accurately is of great significance to ensuring road traffic safety.Traditional pedestrian detection algorithms detect pedestrians based on artificially extracted low-level features.When dealing with deformations such as size,posture,clothing and shelter,the robustness of the algorithms is poor.In recent years,deep learning has been rising in the field of computer vision.Pedestrian detection algorithm based on deep learning technology utilizes complex neural network to extract deep features to be detected in images to represent them,and has good robustness to complex deformation.Although the pedestrian detection algorithm based on deep learning has a good detection effect,it relies on a huge amount of computing resources.Application scenarios such as unmanned driving and advanced driver assistance systems are difficult to provide such huge computing power support.Aiming at this contradiction,this paper studies a lightweight deep learning pedestrian detection algorithm for unmanned driving,which has a high detection accuracy and a small model volume.This is of great significance for the application of deep learning methods to the above scenarios with limited computing resources.The main work and innovation points of this paper are as follows:1.An improved model YOLO-PD based on YOLOV3 is proposed.Through data enhancement and other pre-processing methods,the ability of detecting small targets is improved.By improving the backbone network of YOLOV3 and adding selfattention mechanism,the network has the ability to identify which channel features are more conducive to pedestrian detection,so as to assign higher weight to corresponding features.By introducing the CIOU loss function,the learning ability of the network is improved,so that the network can optimize the situation of more border intersections,thus improving the detection accuracy.2.An improved model based on YOLO-PD is proposed.By introducing deep separable convolutional layer and Crelu activation function,YOLO-PD is improved and a new model YOLO-LPD is obtained.Compared with traditional convolution,deep separable convolution can greatly reduce the number of parameters and calculation amount of convolution layer.CRELU activation function can eliminate the redundant information of convolution kernel of specific convolution layer to achieve the effect of reducing the number of parameters.3.Based on Jetson TX2 development board,a hardware test system is built.Using City Person and Caltech Pedestrian data set,the comparative experiments of YOLOLPD,YOLOV3,YOLOV3-Mobile Net and YOLOV3-Shuffle Net proposed in this paper are carried out.According to their detection accuracy and detection speed as well as the size of their respective models,comprehensive analysis is carried out.Experiments show that the improved algorithm YOLO-PD in this paper has higher accuracy for small target detection and occlusion detection.The YOLO-PD model studied in this paper improves about 3% on the City Person data set compared with the MAP of the original Yolov3 model,and the detection accuracy of the model improves about 5% on the Caltech Pedestrian data set.The volume of the original YOLOV3 model was about 3.78 times that of YOLO-LPD,and the detection speed of YOLOLPD was 2.9 times faster than that of YOLOV3.