3D Pedestrian Detection Based on Pointpillar: SelfAttention-pointpillar

With the rapid development of autonomous Vehicls technology, the detection of surrounding pedestrians, vehicles, Cyclists and other targets by autonomous vehicles is an indispensable technology, especially for pedestrian detection. Therefore, there aremore and more related algorithmms and network models based on target recognition. In recent years, many scholars have stagnated in the process of discovering new algorithms and models and have rarely improved (such as SECOND, PointRCNN, PointPillars, etc.) These classic models, due to the different configuration environments of these model codes and the need to redownload each time you want to run a new model, are very time consuming and energy consuming. In order to solve these difficulties, we chose to use the OpenPCDet target detection framework to improve these models. This framework integrates all the above original object detection models to facilitate us to improve and compare the indicators between the models, and in the comparison of the results of the original model built in the OpenPCDet framework, it is found that the PointPillars modelusing 3D single-stage object detection is the most suitable for autonomous Vehicles. The recognition speed of the original PointPillars for vehicles, pedestrians and other objects can fully meet the use of autonomous Vehicles technology, but the accuracy of object recognition, especially in pedestrian detection, needs to be improved. In this regard, we propose a SelfAttention-PointPillars model. Based on the architecture of the PointPillars model and the idea of self-attention, we use our own pillar amount and modify the original backbone structure into our own self-attention network to improve the accuracy of identifying target pedestrians. We also improved the original L1 loss function into a faster weighted $L 2$ function and we also replaced the activation function with the more efficient LeakyRelu function. Therefore, this paper mainly introduces the OpenPCDet target detection framework design SelfAttention-PointPillars and ensure the recognition speed of the benchmark mode in the KITTI dataset while improving the pedestrian target detection accuracy of bev and 3D from 50 to about 60 and improving the accuracy of the improvement in difficult scenarios is increased by 15 percentage points.