VEHICLE AND PEDESTRIAN DETECTION FOR AUTONOMOUS VEHICLES BASED ON MOBILEVIT

Parfentyev Kirill Viktorovich, Zhang Bohan

Since an autonomous driving information sensing and fusion system needs to identify traffic conditions and various obstacle attributes in a timely and accurate manner, it is especially important to develop an obstacle detection model that achieves both high detection speed and high accuracy. First, a target detection model combining convolutional neural network and Vision Transformer is designed to extract local and global information about vehicles in images. Secondly, an attention-grabbing mechanism module has been introduced to enhance the model's ability to focus on regions; To improve the effect of multi-scale object fusion, double triple interpolation is implemented in the upsampling module. Finally, to ensure real-time performance of the model, the lightweight MobileVit network is used as the model backbone. Experimental results show that AM-Swin Transformer, a MobileVit-based fast vehicle and pedestrian detection model in front of autonomous vehicles proposed in this paper, performs better than other models in terms of vehicle and pedestrian detection accuracy and speed.

Keywords:computer vision, artificial neural networks, object detection, self-driving cars.