Research On End-to-End Autonomous Driving Road Environment Object Detection Algorithm

As a fundamental task in computer vision,Object detection aims at developing novel computational models and methods to improve the recognition accuracy and practical detection speed,and reduce computational overhead.Its challenge depends on different detection tasks involving various objects and difficulties,such as the scale,density,occlusion,and rotation angle of the target.As a crucial technical pillar for the environment perception system for autonomous driving.It seeks to find a highprecision,low-latency,and easy-to-deploy object detector.To address these challenges,we propose a YOLOv5-based improved network model that tackles the road environment object detection task in autonomous driving scenarios.The innovations and specific work are as follows:(1)We present MCS-YOLO.It incorporates the coordinate attention mechanism into the backbone to fuse spatial and cross-channel information,and accurately locate the objects to be detected.It adopts the multi-scale structure to boost the model’s performance on recognizing dense small objects.It integrates the swin transformer into the neck part to improve the network’s global representation ability.(2)We propose a novel DDS-YOLO algorithm.The proposed algorithm uses deformable convolution to adaptively obtain the receptive field of different deformed objects in the road environment and accurately extract the features of complex environment objects,that replaces the network upsampling method with transposed convolution to autonomously learn the optimal upsampling parameters that match the road environment perception task.Also,the proposed algorithm adds a small object detection structure to enhance the model’s sensitivity to dense small objects in the detection task.Aiming at the direction mismatch problem between the real box and the predicted box,the algorithm speeds up the prediction box to converge quickly to the nearby direction by SIoU loss function,effectively improving the training speed and inference accuracy.We conduct ablation experiments and comparative experiments on BDD100 K.The experimental results show that the mean average precision of the MCS-YOLO algorithm reaches 53.6%,the recall rate reaches 48.3%,compared with YOLOv5 s algorithm,that are increased by 4.3% and 3.9% respectively,and the real-time detection speed reaches 55 frames per second.The MCS-YOLO algorithm can effectively improve the detection accuracy and maintain the real-time performance of detection speed for autonomous driving in complex road environment and different weather conditions.The mean average precision of the DDS-YOLO algorithm reaches 55.4%,the recall rate reaches 50.3%,compared with YOLOv5 s algorithm,that are increased by 6.1 percentage points and 5.9 percentage points respectively,the detection speed reaches 71 FPS,and the generated model size is 15.5MB.The DDS-YOLO algorithm is easy to deploy on mobile devices,can effectively improve the detection accuracy and maintain the real-time performance of detection speed for autonomous driving in complex road environment and different weather conditions.