Machine Vision Based Road Pedestrian Detection And Tracking Method

With the increasing number of cars,the problem of road over-saturation and traffic congestion is becoming increasingly prominent,the traffic flow exceeds the road load,and road traffic accidents occur frequently.Intelligent driving assistance can assist drivers according to real-time road conditions,so it has become one of the research hotspots.At present,image processing based on deep learning has become the mainstream direction of visual perception research for intelligent assisted driving.However,the detection effect of existing perceptual algorithms in complex scenes still has a large room for improvement.Therefore,improving the detection accuracy of the target has become an urgent problem for the current pedestrian detection and tracking technology.In order to improve the safety of drivers and road users,this paper studies pedestrian detection and tracking algorithms on traffic roads.Taking pedestrians on the traffic road as the research object,the camera is used to collect pedestrian image information,and the context information is used to complete the detection and tracking of pedestrians through the adaptive weight allocation method,so as to provide reliable information for assisting drivers to avoid obstacles.The specific research contents are as follows:Based on the detection framework of YOLO V4,this paper proposes a pedestrian detection method that utilizes context information,expands receptive field and allocates weight adaptively.First,by enhancing the remote dependency modeling ability of convolutional neural network,the shortcoming of too small target is not easy to extract features.Secondly,aiming at the problem of increasing calculation amount and decreasing detection speed caused by the above steps,a quantization method that can balance accuracy and speed is adopted to improve the reliability of the model without reducing detection speed.Finally,the detection network is tested on the COCO dataset.The experiment shows that the improved detection network can identify pedestrians well,and the detection frame is more suitable for the appearance of pedestrians.Aiming at the disadvantage that the detection frame of the same object changes greatly on different frames due to the different speed of object motion between adjacent frames,this paper uses DeepSORT algorithm and optimizes the Re ID model used by it.To solve the problem that the number of Re ID parameters and the amount of computation increase exponentially with the expansion of the scale of neural network,the lightweight network Shuffle Net V2 is used to replace the original Re ID network,and it is discussed.The experimental results show that the optimized algorithm can reduce the weight size to 1/18 of the original while maintaining the accuracy.Finally,a monocular vision pedestrian detection system is built in the intelligent vehicle dynamic simulation system to verify the proposed algorithm.The video data is collected by the visual simulation system,and the integration of the optimized YOLO V4 model and DeepSORT is realized through the collaborative work of related hardware modules and software systems,and the real-time pedestrian detection and tracking results are finally displayed in the display.The experimental results show that the proposed pedestrian detection and tracking algorithm achieves satisfactory results in the experiment.