Design And Implementation Of Pedestrian Detection System Based On Regions With Convolutional Neural Networks

Pedestrian detection has important value in application fields such as video surveillance and automatic driving,and is one of the research hotspots in object detection.Due to the particularity of the application scenario,the pedestrian detection algorithm needs to meet some special requirements: 1.Under the premise of ensuring the detection accuracy,the whole detection process needs to be as fast as possible to meet the real-time requirements;2.Small-scale pedestrians in the pedestrian detection dataset occupies a large proportion,and the detection accuracy of this part of the pedestrian needs to be further improved,and the detection recall rate must be guaranteed.Regions with convolutional neural networks is one of the most successful solutions in general object detection.It has become a typical object detection framework and is widely used in pedestrian detection.The thesis studies the above problems based on regions with convolutional neural networks.The main work is as follows:(1)The traditional convolutional neural networks pedestrian detection algorithm can not effectively deal with continuous image detection.The thesis proposes a regions with convolutional neural networks pedestrian detection algorithm based on historical information.The algorithm uses the detection result of the previous frame as the reference information of the region proposal of the current frame,and determines whether there is still a pedestrian near the position of the pedestrian detected in the previous frame,and according to the difference value of the gray value between the current frame and the previous frame,The convolution feature of the current image is filtered to narrow the search area when the area region proposal network uses the sliding window detection,thereby improving the quality and extraction speed of the region proposal.The test results on the pedestrian detection dataset show that the method effectively improves the speed of continuous pedestrian detection.(2)To solve the problem of low accuracy of small-scale pedestrian detection,the thesis designs the feature aware regions with convolutional neural networks pedestrian detection algorithm.In the feature map extracted by the convolutional neural networks,the resolution of the convolution feature of the small-scale pedestrian is low,resulting in poor detection results of the subsequent detection network.The thesis uses the feature-aware algorithm to select the appropriate convolutional layer output from the convolutional neural network as the input of the detection network according to the size of the region proposal.This process is equivalent to dynamically adjusting the depth of the convolutional neural networks,making the volume of pedestrians of different scales.The product features both semantics and resolution,which in turn improves the accuracy of small-scale object detection.(3)Using the above-mentioned optimized pedestrian detection scheme,the thesis implements a pedestrian detection system based on the Raspberry Pi design.The TensorFlow Lite runtime environment was built on the Raspberry Pi,and the TensorFlow model conversion tool was used to convert the trained pedestrian detection model into a format suitable for running on TensorFlow Lite for model deployment.The experimental results in the actual scene show that the pedestrian detection system can effectively detect pedestrians of different scales.