Light Weight Real-time Network Design For Automatic Driving Target Detection

The automatic driving controlling system is an autonomous decision-making system,which mainly processes the observation data from the on-board radar and visual detection sensors.At present,the target detection technology based on vehicle radar is relatively mature.In order to improve the entire set of automatic driving systems,the requirements for the visual detection performance of the convolutional neural network are getting higher and higher.In the visual inspection of convolutional neural network,there have been many high-precision target detection networks,such as YOLO,but such networks with high structure complexity cannot be effectively embedded in small processors for real-time use effectively.In the vision system of automatic driving,the processor of the camera is small with limited calculation speed.In order to satisfy that the target detection network can be embedded in a small processor and effectively complete the target detection task,the target detection network structure needs to continuously miniaturize the detection model and reduce its weight as much as possible while still ensuring a certain accuracy.The thesis focuses on the construction of a lightweight two-stage target detection network structure,and achieves its good performance in the real-time detection of targets on the mobile terminal.The thesis achieves real-time detection of targets by building a lightweight two-stage target detection network structure,and mainly completes the following tasks:1)The algorithm and training process of target detection network are studied,and the current popular target detection network structure is analyzed.The performance limitations of these algorithms are summarized;2)Lightweight models are designed from three aspects,the efficiency,accuracy and stability of the lightweight network.Saliency grouping convolution is designed in terms of efficiency.This model divides the feature map into multiple subspaces,reduces the amount of calculation and enhances the fluidity of effective information among channels.Considering the improvement of accuracy in the case of low complexity,the design of a multi-layer group perceptron suitable for lightweight classification networks can replace the deep separable convolution operation in the specified location.In view of the unstable convergence performance of the SGD optimizer during the lightweight model training process,the SGD optimizer based on horizontal standardization is designed.The accuracy and loss function fulfill the expected requirements during the training process;3)A lightweight real-time target detection network which is based on two-stage is built.Our network achieves not only excellent real-time performance of a single-stage detection network and the high accuracy of a two-stage network,but also rich location information of the shallow network and high degree of discrimination of deep network features.The RPN network is introduced into our network,and the information of the foreground and background is combined to optimize the feature distribution.As a result,an efficient and low computational load target detection network is obtained;4)Special pictures,such as pedestrian occlusion in the VOC data set,are expanded,and the performance of the combination of our target detection network and tracking algorithm is verified.At last,Our target detection network is embeded on the mobile terminal to complete real-time target detection.By comparing with the performance indicators of existing target detection networks,the two-stage lightweight real-time target detection network in this paper can achieve better detection performance on the mobile phone.