Research On Pedestrian Detection Technology Based On On-board Vision Systems

In recent years,with the development of artificial intelligence and the improvement of hardware computing capabilities,the pedestrian detection algorithm has also been greatly improved in performance and applied in some practical scenarios.However,due to the large computational resources required for deep learning,the pedestrian detection algorithm based on deep learning can not be well applied in scenes with limited computing resources and high real-time requirements,such as automatic driving and mobile robots,etc.In this thesis,the SSD(Single Shot Multi Box Detector)objective detection algorithm is used for pedestrian detection,and the research on the lightweight of the proposed algorithm is carried out around the vehicle camera scene.The main research contents are as follows:(1)In order to solve the problem of large amount of computation and long timeconsuming of feature extraction network in SSD target detection algorithm,this thesis replaces the original feature extraction network with the Mobile Net lightweight feature extraction network,which effectively reduces the computation amount and the number of parameters in the algorithm.Follow the idea of using multi-scale feature maps for objective detection in SSD,Considering the condition that the pedestrian scale varies greatly and small scale targets account for a large proportion in the scene of the car camera,this thesis uses seven feature maps with different scales to pedestrian target detection.The detection accuracy for small-scale targets has been improved by 1.35%.(2)Aiming at the pedestrian target detection task under the scene of vehicle camera,this thesis statistically analyzes the aspect ratio of pedestrian target in the dataset and readjusts the aspect ratio and the number of the default detection box in the algorithm.Three default detection boxes with different aspect ratios are designed in this thesis,and the number of default detection boxes in different feature maps is adjusted accordingly.Finally,the number of default boxes generated is reduced by half compared with corresponding default detection boxes generated by the native SSD.(3)To solve the problem of limited floating-point computational ability of embedded platform,this thesis designs a new activation function based on the common one to limit the output range,alleviating the computational difficulty and increasing the response to negative input.In this thesis,the above-mentioned network design and optimization are experimentally verified,and the proposed algorithm is compared with two main pedestrian detection algorithms.At the same time,the robustness of the proposed algorithm is tested under different weather scenarios.The experimental results show that the algorithm in this thesis is robust and has better detection effect with recall 66.73%.It basically meets the real-time requirements on the NVIDIA Jetson TX2 embedded platform.