Visual Attention Estimation Algorithm and Dynamic Neural Network Based Object Detection for Intelligent Vehicles

Vision-based object detection for intelligent vehicles differs from regular object detection tasks. It necessitates an attention region where objects have a critical impact on the vehicle’s driving behavior. This region is influenced by many factors, especially the ego-vehicle’s driving state and the surrounding traffic participants. Existing approaches uniformly process the entire image, leading to significant computation redundancy. The objective of this study is to accurately and efficiently estimate this region and to enhance detection capability within it, thereby improving the detection performance of critical objects. Initially, an attention region estimation method is proposed. This method conceptualizes the estimation of the attention region as a mass center computation, taking into account the ego-vehicle’s driving state and the surrounding traffic participants. Subsequently, an object detection method based on dynamic neural network is proposed. This method adaptively allocates computational resources according to the attention region. It employs an uneven downsampling method during preprocessing and introduces a dynamic block into the network architecture, which is designed to optimize feature representation within the attention region, thereby enhancing detection performance. To verify the effectiveness of the proposed method, we established a real-vehicle dataset, in which the critical objects are annotated by the drivers. Compared to the baseline model, the mean average precision (mAP) of critical objects has increased by 3.6%, while the average computation has decreased by 23.6%. The proposed method offers an efficient and accurate approach to object detection for intelligent vehicles, with the potential to enhance driving safety and reduce computational demands.