Research On Test Technology Of Lidar And Camera Target Recognition System For Automatic Driving

Target recognition technology is one of the research hotspots in AI perception technology.It is widely used in automatic driving,face recognition,robotics,military and other fields.Because the detection objects,evaluation indicators and software algorithms of the target recognition system are different,it brings trouble to the selection of the target recognition system in the actual application process.Therefore,the objective and systematic evaluation of the target recognition system becomes the key to distinguish the performance of different target recognition systems.Aiming at the target recognition system for lidar and depth camera for night automatic driving,this paper discusses their key measurement technologies from the two aspects of sensor detection ability and algorithm recognition ability,and designs the corresponding evaluation methods.The main contents are as follows:（1）The night-time scene is analyzed,and two parameters of illuminance and Bidirectional Reflectance Distribution Function（BRDF）are proposed to simulate the night-time scene.The YOLO model was trained for this scene,and the construction of the target recognition system was completed.Sensitivity index,evaluation method based on minimum distance classifier and F1-score are used to evaluate the recognition ability of sensors and algorithms.Compared with traditional methods,the detection ability of different sensors can be quantitatively evaluated.（2）The BRDF standard body is made to simulate night objects,and the sensitivity index is used to evaluate the target detection ability in hardware and distance detection ability of lidar under night-time environment.The experimental results show that the lidar can detect the target for the standard volume of 0.0068 sr^-1～0.34 sr^-1 BRDF in the illumination range of 0～5 lx at a distance of 1m.（3）The target detection of RGB camera and structured light camera at night scene is realized by using BRDF grayscale standard body and sensitivity index.At the same time,the evaluation method based on the minimum distance classifier realizes the lateral evaluation of the detection ability of the three sensors.Secondly,the BRDF standard arrow based on the ink screen is made,and the target recognition ability of the YOLO algorithm is evaluated by the F1-score evaluation index.The experimental results show that the RGB camera can detect samples with the BRDF of 0.0571 sr^-1under 0.51 lx,and can detect samples with the BRDF of 0.0291 sr^-1 under 0.98 lx.Compared with the traditional evaluation method,the F1-score evaluation can guide the practical application and improvement direction of the target recognition system.